Cyclic matching pursuits with multiscale time-frequency dictionaries

We generalize cyclic matching pursuit (CMP), propose an orthogonal variant, and examine their performance using multiscale time-frequency dictionaries in the sparse approximation of signals. Overall, we find that the cyclic approach of CMP produces signal models that have a much lower approximation error than existing greedy iterative descent methods such as matching pursuit (MP), and are competitive with models found using orthogonal MP (OMP), and orthogonal least squares (OLS). This implies that CMP is a strong alternative to the more computationally complex approaches of OMP and OLS for modeling high-dimensional signals.