Digital predistortion of radio frequency power amplifiers using Kautz-Volterra model

Introduction: Power amplifiers (PAs) with high power efficiency are inherently nonlinear (NL). The requirements on linearity have increased since wideband signals with high peak-to-average ratios are used more in, e.g. mobile systems. Linearisation is, thus, necessary for high power efficiency of PAs. Digital predistortion (DPD) has attracted much interest since it is a cost-effective linearisation technique. In DPD the signal is distorted in the digital domain to compensate for the PA’s signal distortion. In model based DPD an approximate inverse of the PA’s transfer function is used. The inverse can be complicated also for relatively simple NL systems [1]. PAs often have both ‘linear’ and ‘NL memory effects’ [2] and an exact inverse could not be found even if a PA’s direct transfer function were known. The inverse model is also NL with memory, but has, in many cases, longer memory than the direct model of the same system [1]. Neural networks can be used for approximating NL systems with memory and have been used for DPD [3]. A Volterra model, which can be used for any causal stable NL system with fading memory, could in principle also be used. The high number of parameters and the computational complexity make both neural networks and a Volterra model unsuitable in practice. Models that are special cases of the Volterra model are, thus, used [4, 5]. These models, sometimes referred to as memory polynomials, may give good results when used in DPD algorithms, although they do not have the general properties of a Volterra model. In this Letter we report the use of an odd order Kautz-Volterra (KV) model as a DPD algorithm. The model uses orthonormal basis functions, so-called Kautz functions [6]. The KV model has the same general properties as the Volterra model, but the number of parameters can be significantly lower. A Volterra model is based on FIR filters; a KV model uses IIR filters and can in practice be used for systems with longer memory effects [6]. KV models have, to our knowledge, previously not been used for inverse models, such as a DPD algorithm. The KV model has the advantage over memory polynomials that it could be used for any NL system with fading memory. In the KV model the basis functions of each NL order are orthonormal. DPD algorithms using orthogonal polynomials have been reported. In these the orthogonality depends on the signals probability density function [7]. Loosely speaking one could say that the KV model is orthogonal in memory and orthogonal polynomials in amplitude.