Abstract:A covariance-free method of computing kernel principal components is proposed. First, a matrix, called Gram-power matrix, is constructed with the original Gram matrix. It is proven by the theorem of linear algebra that the eigenvectors of newly constructed matrix are the same as those of the Gram matrix. Therefore, each column of the Gram matrix can be treated as the input sample for the iterative algorithm. Thus, the kernel principle components can be iteratively computed without the eigen-decomposition. The space complexity of the proposed method is only O(m), and the time complexity is reduced to O(pkm). The effectiveness of the proposed method is validated by experimental results. More importantly, it still can be used even if traditional eigen-decomposition technique cannot be applied when faced with the extremely large-scale data set.