1: \begin{abstract}

2:   We observe $(X_i,Y_i)_{i=1}^n$ where the $Y_i$'s are real valued

3:   outputs and the $X_i$'s are $m\times T$ matrices. We observe a new

4:   entry $X$ and we want to predict the output $Y$ associated with it.

5:   We focus on the high-dimensional setting, where $m T \gg n$. This

6:   includes the matrix completion problem with noise, as well as other

7:   problems. We consider linear prediction procedures based on

8:   different penalizations, involving a mixture of several norms: the

9:   nuclear norm, the Frobenius norm and the $\ell_1$-norm. For these

10:   procedures, we prove sharp oracle inequalities, using a statistical

11:   learning theory point of view. A surprising fact in our results is

12:   that the rates of convergence do not depend on $m$ and $T$ directly.

13:   The analysis is conducted without the usually considered incoherency

14:   condition on the unknown matrix or restricted isometry condition on

15:   the sampling operator.  Moreover, our results are the first to give

16:   for this problem an analysis of penalization (such nuclear norm

17:   penalization) as a regularization algorithm: our oracle inequalities

18:   prove that these procedures have a prediction accuracy close to the

19:   deterministic oracle one, given that the reguralization

20:   parameters are well-chosen. \\

21: 

22:   \noindent%

23:   \emph{Keywords.} High dimensional matrix ; Matrix completion ;

24:   Oracle inequalities ; Schatten norms ; Nuclear norm ; Empirical risk

25:   minimization~; Empirical process theory~; Sparsity

26: \end{abstract}

27: