1: \begin{abstract}

2: We propose a framework to shrink a user-specified characteristic of a precision matrix estimator that is needed to fit a predictive model. 

3: Estimators in our framework minimize

4: the Gaussian negative log-likelihood plus an

5: $L_1$ penalty on a linear or affine function evaluated at the optimization variable corresponding to the

6: precision matrix. We establish convergence rate 

7: bounds for these estimators and we propose an alternating direction method of 

8: multipliers algorithm for their computation. 

9: Our simulation studies show that our estimators 

10: can perform better than competitors when they are used to fit predictive models.  In particular, we illustrate cases where 

11: our precision matrix estimators perform worse at estimating the population precision matrix while performing better at prediction.

12: \end{abstract}

13: