Gradient pursuit for non-linear sparse signal modelling

In this paper the linear sparse signal model is extended to allow more general, non-linear relationships and more general measures of approximation error. A greedy gradient based strategy is presented to estimate the sparse coefficients. This algorithm can be understood as a generalisation of the recently introduced Gradient Pursuit framework. Using the presented approach with the traditional linear model but with a different cost function is shown to outperform OMP in terms of recovery of the original sparse coefficients. A second set of experiments then shows that for the non-linear model studied and for highly sparse signals, recovery is still possible in at least a percentage of cases.