Natural Gradient-based Adaptive Algorithms for Sparse Underwater Acoustic Channel Identification

Natural-gradient (NG) adaptive algorithms are known to be superior to stochasticgradient (SG) algorithms when the channel to be identified exhibits a known Riemannian structure. In sparse channel identification, for example, the improved-proportionate normalized least-mean-square (IPNLMS) is a well known NG algorithm which outperforms the classical SG normalized least-mean square (NLMS). Apart from Riemannian geometry, the L0 norm is an alternative way to describe channel sparseness. In this work, a new algorithm is proposed for sparse underwater acoustic (UWA) channel estimation by incorporating the Riemannian metric of the IPNLMS and an appropriately defined L0 norm of the channel taps into the same cost function. To cope with the high Doppler frequencies often encountered in UWA channels, carrier-phase estimation is incorporated into the algorithm. Based on data recorded over a sparse acoustic link, the superior performance of the proposed algorithm to existing NG and SG algorithms is validated.