Review for NeurIPS paper: Efficient Learning of Discrete Graphical Models

Weaknesses: It seems like the main difference compared to [17] is to use general basic function in equation (11), instead of focusing on Ising model as in [17]. Other than this, the structure of the paper is similar to [17], e.g. the Local learnability condition corresponds to the restricted strong convexity, e.g. the gradient concentration. The value \hat\gamma in (5) is claimed as the prior information on the parameter (line 271). Based on my understanding it plays a similar role as \lambda in [17]. Is there any particular reason to switch from L1 regularization in [17] to L1 constrained optimization here?