Retrospective Mutiple Change-Point Estimation with Kernels

This contribution proposes an extension of the classic dynamic programming algorithm for detecting jumps in noisily observed piecewise-constant signals. The proposed algorithm operates (virtually) in a reproducing kernel Hilbert space through the use of an arbitrary kernel mapping. The resulting approach provides a computationally efficient an versatile tool for segmenting complex signals whose structure is not appropriately captured by standard parametric models.