Giheon Seong

Standard PCA is strictly suboptimal for high-dimensional representation recovery: shrinking sample eigenvectors toward subspace targets that borrow strength across complementary spiked components guarantees dominance over empirical eigenspaces without degrading under uninformative priors.