Learning rates of regularized regression on the unit sphere (Q365838)

This paper studies least square regularized regression learning algorithms with spherical harmonics kernels on the unit sphere. The learning error is decomposed into sample error and regularized error. To estimate the regularized error, the best approximation operator and the Jackson-type inequality are introduced. The other interesting point is that the dimension of hypothesis space charges with the size of samples. Satisfying error bound and learning rate are deduced by the covering number technique.