Concentration bounds and asympotic distribution for the empirical spectral projectors of sample covariance operators

Speaker(s): 
Karim Lounici (Georgia Institute of Technology, Atlanta)
Date: 
Wednesday, December 9, 2015 - 10:00am
Location: 
Paul-Drude-Institut für Festkörperelektronik, Hausvogteiplatz 5-7, 10117 Berlin, EG Raum 007

Let $X,X_1,\dots, X_n$ be i.i.d. Gaussian random variables in a separable Hilbert space ${\mathbb H}$ with zero mean and covariance operator $\Sigma={\mathbb E}(X\otimes X),$ and let $\hat \Sigma:=n^{-1}\sum_{j=1}^n (X_j\otimes X_j)$ be the sample (empirical) covariance operator based on $(X_1,\dots, X_n).$ Denote by $P_r$ the spectral projector of $\Sigma$ corresponding to its $r$-th eigenvalue $\mu_r$ and by $\hat P_r$ the empirical counterpart of $P_r.$ We derive tight bounds on $$ \sup_{x\in {\mathbb R}} \left|{\mathbb P}\left\{\frac{\|\hat P_r-P_r\|_2^2-{\mathbb E}\|\hat P_r-P_r\|_2^2}{{\rm Var}^{1/2}(\|\hat P_r-P_r\|_2^2)}\leq x\right\}-\Phi(x)\right|, $$ where $\|\cdot\|_2$ denotes the Hilbert-Schmidt norm and $\Phi$ is the standard normal distribution function. These bounds depend on the so called effective rank of $\Sigma$ defined as ${\bf r}(\Sigma)=\frac{{\rm tr}(\Sigma)}{\|\Sigma\|_{\infty}},$ where ${\rm tr}(\Sigma)$ is the trace of $\Sigma$ and $\|\Sigma\|_{\infty}$ is its operator norm, as well as another parameter characterizing the size of ${\rm Var}(\|\hat P_r-P_r\|_2^2).$ For an eigenvalue $\mu_r$ of $\Sigma$ of multiplicity $1$ with associated eigenvector $\theta_r$, we derive new properties about the bias of the sample covariance eigenvector as an estimator of $\theta_r$. As a consequence, we suggest a new simple estimator of $\theta_r$ with decreased bias and we derive a concentration bound on the $l_{\infinity}$-norm of the deviation between the estimator and $\theta_r$. This result may be of interest to perform variable selection. (Joint work with Vladimir Koltchinskii)