1: \begin{abstract}

2: We propose a scalable, efficient and statistically motivated computational framework for Graphical Lasso \citep{FHT2007a} --- 

3: a covariance regularization framework that has received significant attention in the statistics community over the past few years. 

4: Existing algorithms have trouble in scaling to dimensions larger than a thousand. 

5: Our proposal significantly enhances the state-of-the-art for such moderate sized problems  

6: and gracefully scales to larger problems where other algorithms become practically infeasible. 

7: This requires a few key new ideas. We operate on the primal problem and use a subtle variation of block-coordinate-methods which drastically reduces the computational 

8: complexity by orders of magnitude. We provide rigorous theoretical guarantees on the convergence and complexity of our algorithm and 

9: demonstrate the effectiveness of our proposal via experiments. 

10: We believe that our framework extends the applicability of Graphical Lasso to large-scale modern applications like bioinformatics, collaborative filtering and social networks, among others. 

11: \end{abstract}