A Doubly Orthogonal Matching Pursuit Algorithm for Sparse Predistortion of Power Amplifiers

Abstract

This letter presents a new method for the digital predistortion of power amplifiers (PAs) based on sparse behavioral models. The Gram-Schmidt orthogonalization is synergistically integrated into the orthogonal matching pursuit algorithm to decorrelate the selected model regressors against the components still to be selected. Experiments conducted on a test bench based on a GaN PA driven by a 15-MHz orthogonal frequency division multiplexing signal were conducted in order to validate the algorithm. Experimental results in a digital predistortion application and a comparison with other state-of-the-art algorithms highlight the enhancement of its pruning capabilities, reducing the number of coefficients while maintaining the performance.

Sparse Identification of Volterra Models for Power Amplifiers Without Pseudoinverse Computation

“…Considering that the matrices defined earlier are sized by the number of model components, which is generally much lower than the number of samples in the regressors, the resulting algorithm shows a computational complexity that is fixed for every iteration compared with that of the previous versions of the algorithm [11], [13], [21] that required increasing computations with the number of selected coefficients. Sections III-A-III-F overview the steps of this contribution.…”

“…where O stands for the quantity of real-number multiplications at iteration k. It can be observed that the original DOMP [11] exhibits a complexity dominated by the cubic of the iteration number that results of the m × k pseudoinverse calculation at each iteration; in the SSPI DOMP [21], this relationship is reduced to quadratic dependence, and in the novel RC-DOMP, there is no dependence with the iteration number. Considering that m/n is the number of samples per regressor, it is straightforward to conclude that the RC-DOMP outperforms the rest of the algorithms under comparison.…”

“…Within the family of greedy pursuits, which are characterized for selecting regressors by hard decision iteratively, the orthogonal matching pursuit (OMP) was applied to PA DPD in [5], followed by a reduced complexity version named by the authors as a simplified sparse parameter identification (SSPI) algorithm [6] that turned the calculation of the pseudoinverse matrix at each step into an iterative fashion. The modified Gram-Schmidt (MGS) technique [7], also known as orthogonal least squares (OLS) [8]- [10], and OMP were fused into the doubly OMP (DOMP) [11], which was shown to overcome the difficulties of choosing the correct regressors under a highly correlated basis such as the Volterra series. The main issue of the use of MGS against OMP is the significant increase in computational complexity [12], although the latter has shown to be outperformed by the DOMP in the pruning of both PA and DPD models [13].…”

“…The DOMP algorithm in [11] presents a high computational complexity due to the pseudoinverse operation and the Kronecker product needed for the orthogonalization. In [21], an SSPI DOMP was contributed, which showed a lower computational complexity by transforming the pseudoinverse calculation in a recursive operation that depends on the result of the previous iteration.…”

Comparative Analysis of Greedy Pursuits for the Order Reduction of Wideband Digital Predistorters

“…The orthogonal matching pursuit (OMP) was first applied to PA DPD in [16], followed by a reduced-complexity version of the algorithm in [17]. To overcome the high correlation that appears in Volterra series, the doubly OMP (DOMP) was designed [18] as well as its low-complexity variant [19].…”