1: \begin{abstract}

2:   Boob et al. \cite{flowless} described an iterative peeling algorithm

3:   called \gplus for the Densest Subgraph Problem (\dsg) and

4:   conjectured that it converges to an optimum solution.  Chekuri,

5:   Qaunrud and Torres \cite{chandra-soda} extended the algorithm to

6:   general supermodular density problems (of which DSG is a special

7:   case) and proved that the resulting algorithm

8:   \textsc{Super-Greedy++} (and hence also \textsc{Greedy++})

9:   converges. In this paper we revisit the convergence proof and

10:   provide a different perspective.  This is done via a connection to

11:   Fujishige's quadratic program for finding a lexicographically optimal base

12:   in a (contra) polymatroid \cite{fujishige}, and a noisy version of the Frank-Wolfe

13:   method from convex optimization \cite{FW-56,pmlr-v28-jaggi13}.  This gives us a simpler convergence

14:   proof, and also shows a stronger property that \textsc{Super-Greedy++}

15:   converges to the optimal dense decomposition vector, answering a

16:   question raised in Harb et al. \cite{farouk-neurips}. A second

17:   contribution of the paper is to understand Thorup's work on ideal tree packing and 

18:   greedy tree packing \cite{Thorup07,Thorup08} via

19:   the Frank-Wolfe algorithm applied to find a lexicographically optimum

20:   base in the graphic matroid.  This yields a simpler and transparent proof. The two results appear disparate but are unified via Fujishige's result and convex optimization. 

21: \end{abstract}

22: