Testing for Principal Component Directions under Weak Identifiability

TL;DR

This paper examines the limitations of classical tests for principal component directions under weak identifiability, proposing a more robust testing approach that maintains correct size and power in challenging asymptotic scenarios.

Contribution

It introduces a new understanding of the behavior of likelihood ratio and Le Cam optimal tests under weak eigenvalue separation, advocating for the latter in such cases.

Findings

Le Cam optimal test maintains nominal level under weak identifiability.

Likelihood ratio test overrejects when eigenvalues are close.

The new test is robust and retains power in challenging asymptotic regimes.

Abstract

We consider the problem of testing, on the basis of a $p$-variate Gaussian random sample, the null hypothesis ${\cal H}_0: {\pmb \theta}_1= {\pmb \theta}_1^0$ against the alternative ${\cal H}_1: {\pmb \theta}_1 \neq {\pmb \theta}_1^0$, where ${\pmb \theta}_1$ is the "first" eigenvector of the underlying covariance matrix and ${\pmb \theta}_1^0$ is a fixed unit $p$-vector. In the classical setup where eigenvalues $\lambda_1>\lambda_2\geq \ldots\geq \lambda_p$ are fixed, the Anderson (1963) likelihood ratio test (LRT) and the Hallin, Paindaveine and Verdebout (2010) Le Cam optimal test for this problem are asymptotically equivalent under the null hypothesis, hence also under sequences of contiguous alternatives. We show that this equivalence does not survive asymptotic scenarios where $\lambda_{n1}/\lambda_{n2}=1+O(r_n)$ with $r_n=O(1/\sqrt{n})$. For such scenarios, the Le Cam optimal test still asymptotically meets the nominal level constraint, whereas the LRT severely overrejects the null hypothesis. Consequently, the former test should be favored over the latter one whenever the two largest sample eigenvalues are close to each other. By relying on the Le Cam's asymptotic theory of statistical experiments, we study the non-null and optimality properties of the Le Cam optimal test in the aforementioned asymptotic scenarios and show that the null robustness of this test is not obtained at the expense of power. Our asymptotic investigation is extensive in the sense that it allows $r_n$ to converge to zero at an arbitrary rate. While we restrict to single-spiked spectra of the form $\lambda_{n1}>\lambda_{n2}=\ldots=\lambda_{np}$ to make our results as striking as possible, we extend our results to the more general elliptical case. Finally, we present an illustrative real data example.