Theory of Matching Pursuit in Kernel Defined Feature Spaces

We analyse matching pursuit for kernel principal components analysis (KPCA) by proving that the sparse subspace it produces is a sample compression scheme. We show that this bound is tighter than the