1: \begin{abstract} \

2: This paper deals with the trace regression model where $n$ entries

3: or linear combinations of entries of an unknown $m_1\times m_2$

4: matrix $A_0$ corrupted by noise are observed. We propose a new

5: nuclear-norm penalized estimator of $A_0$ and establish a general

6: sharp oracle inequality for this estimator for arbitrary values of

7: $n,m_1,m_2$ under the condition of isometry in expectation. Then

8: this method is applied to the matrix completion problem. In this

9: case, the estimator admits a simple explicit form and we prove that

10: it satisfies oracle inequalities with faster rates of convergence

11: than in the previous works. They are valid, in particular, in the

12: high-dimensional setting $m_1m_2\gg n$. We show that the obtained

13: rates are optimal up to logarithmic factors in a minimax sense and

14: also derive, for any fixed matrix $A_0$, a non-minimax lower bound

15: on the rate of convergence  of our estimator, which coincides with

16: the upper bound up to a constant factor. Finally, we show that our

17: procedure provides an exact recovery of the rank of $A_0$ with

18: probability close to 1. We also discuss the statistical learning

19: setting where there is no underlying model determined by $A_0$ and

20: the aim is to find the best trace regression model approximating the

21: data. As a by-product, we show that, under the Restricted Eigenvalue

22: condition, the usual vector Lasso estimator satisfies a sharp oracle

23: inequality (i.e., an oracle inequality with leading constant 1).

24: \end{abstract}