1: \begin{abstract}

2: This paper considers estimation of sparse covariance matrices and

3: establishes the

4: optimal rate of convergence under a range of matrix operator norm and

5: Bregman divergence losses. A major focus is on the

6: derivation of a rate sharp minimax lower bound. The problem exhibits new

7: features that are significantly different from those that occur in the

8: conventional nonparametric function estimation problems. Standard techniques

9: fail to yield good results, and new tools are thus needed.

10: 

11: We first develop a lower bound technique that is

12: particularly well suited for treating ``two-directional'' problems such as

13: estimating sparse covariance matrices. The result can be viewed as a

14: generalization of Le Cam's method in one direction and Assouad's Lemma in

15: another. This lower bound technique is of independent interest and can be

16: used for other matrix estimation problems.

17: 

18: We then establish a rate sharp minimax lower bound for estimating sparse

19: covariance matrices under the spectral norm by applying the general lower

20: bound technique. A thresholding estimator is shown to attain the optimal

21: rate of convergence under the spectral norm. The results are then

22: extended to the general matrix $\ell_w$ operator norms for $1\le w\le

23: \infty$. In addition, we give a unified result on the minimax rate of

24: convergence for sparse covariance matrix estimation under a class of

25: Bregman divergence losses.

26: \end{abstract}