Optimality and sub-optimality of PCA. I: Spiked random matrix models (Q1800806)

The main objective of this paper is to address the following question: ``Can any statistical procedure detect the presence of a prominent eigenvector below the threshold where PCA succeeds?'' To answer this question, it is shown that in the Gaussian Wigner model the spectral threshold \(\lambda =1\) is optimal for priors such as the uniform prior on the unit sphere, the i.i.d. Rademacher prior and any prior with a sufficient sub-Gaussian bound. In the general Wigner model, the authors show that the spectral threshold is never optimal. It is shown that when \(x\) is i.i.d. Rademacher distributed the spectral threshold is only sometimes statistically optimal.