1: \begin{abstract}

2: In the convolution model $Z_i=X_i+ \varepsilon_i$, we give a model

3: selection procedure to estimate the density of the unobserved

4: variables $(X_i)_{1 \leq i \leq n}$, when the sequence $(X_i)_{i

5: \geq 1}$ is strictly stationary but not necessarily independent.

6: This procedure depends on wether the density of $\varepsilon_i$ is

7: super smooth or ordinary smooth. The rates of convergence of the

8: penalized contrast estimators are the same as in the independent

9: framework, and are minimax over most classes of regularity on

10: ${\mathbb R}$. Our results apply to mixing sequences, but also to

11: many other dependent sequences. When the errors are super smooth,

12: the condition on the dependence coefficients is the minimal

13: condition of that type ensuring that the sequence $(X_i)_{i \geq 1}$

14: is not a long-memory process.

15: \end{abstract}

16: