The paper makes an interesting and seemingly deep connection between a well-known algorithm (OSMD) and "information ratio", a notion that arises in the analysis of some online learning algorithms. This connection can potentially be used as a tool/technique in the analysis of online learning algorithms. The paper presents three applications, so we know that the new tool can actually be used. However, these applications obtain only very marginal improvements over the prior work. The paper is all right as is, but it would be much more impressive if the applications obtained substantial improvements and/or simplifications. There was a substantial discussion among the reviewers and the AC.