降维(比如PCA或者random projection) -> KDE(kernel density estimation)来估算密度 -> KL divergence