VC: High Dimensional PCA and Kernel PCA
This post assumes you already know what PCA is. If you don’t please check my previous post.
n < p case (High dimensional PCA)
If the feature space is bigger than the number of data points, our rank is determined by n-1 because it is centred by the mean. Therefore, the degree of freedom is reduced, we can calculate the last one by the previous n-1 data points because of the mean.
We will use one fact to efficiently calculate the high dimensional PCA. All eigenvectors of S are in the span of z(the centred vector of the original data).
This proof starts with the eigenvalue decomposition of the scatter matrix. We can think of the inner product of S and u as the sum of vector inner products. We got u is some scalar product of z. The scalar coefficient consists of eigenvalue, eigenvector, and z transpose. We can reduce the computation with this fact from p x p eigenvalue problem to n x n eigenvalue problem.
We start with the eigenvalue decomposition of the scatter matrix S and it becomes the eigenvalue decomposition of matrix K because its vector product is different, it is from the different…