1: \begin{abstract}

2: We propose dimension reduction methods for sparse, high-dimen\-sional

3: multivariate response regression models. Both the number of responses

4: and that of the predictors may exceed the sample size. Sometimes viewed

5: as complementary, predictor selection and rank reduction are the most

6: popular strategies for obtaining lower-dimen\-sional approximations of

7: the parameter matrix in such models. We show in this article that

8: important gains in prediction accuracy can be obtained by considering

9: them jointly. We motivate a new class of sparse multivariate regression

10: models, in which the coefficient matrix has low rank \textit{and} zero

11: rows or can be well approximated by such a matrix. Next, we introduce

12: estimators that are based on penalized least squares, with novel

13: penalties that impose simultaneous row and rank restrictions on the

14: coefficient matrix. We prove that these estimators indeed adapt to the

15: unknown matrix sparsity and have fast rates of convergence. We support

16: our theoretical results with an extensive simulation study and two data

17: analyses.

18: \end{abstract}