# Separate all leaves of a weighted tree with minimum weight cuts

This is part of a larger problem, which I believe I have reduced to this. Given a tree $T$ having positive edge weights, and $k$ leaves (nodes which have exactly one connected node), I need to delete some edges in the tree so that no two leaves in the original tree are connected (by a path) in the newly formed forest (of trees). The total sum of the weights of the deleted edges needs to be minimized.

My understanding is that atleast $k-1$ edges need to be deleted to separate out all the $k$ leaves. Any more deletions will unnecessarily increase the total cost. Thus, we need to perform exactly $k-1$ deletions.

My hypothesis: For every pair of leaf nodes $l_i$ and $l_j$, find the edge with the minimum weight in the (unique) path from $l_i$ to $l_j$. The $k-1$ least weight edges from this set of edges need to be deleted. This will minimize the sum of weights of the edges to be deleted in order to disconnect all leaves from each other.

I am unable to prove or disprove this hypothesis. Can someone please prove the correctness of this hypothesis, or give a counter-example along with the correct algorithm to solve this problem? If this is indeed correct, is there a faster way (asymptotic complexity wise) to solve this problem? This approach will take $\Theta({k^2})$ time. Thanks in advance!

• think about the problem recursively. Base case: what happens for a sub-tree with height 1? (all nodes but the root are leaves). – Joe Oct 8 '12 at 5:40
• @Joe In that case, the root becomes the LCA (Least Common Ancestor) for all the leaves, so among all paths (of height 1) that lead to and from the root, $k-1$ will have to be deleted. There would be $k$ such paths, so we delete all but the most expensive one. How do I extend the recursion next? By considering a tree with atleast one leaf with a depth more than one? Or a tree with atleast two leaves with an LCA at a height more than 1 from each of them? – Paresh Oct 8 '12 at 13:30
• You delete all but the least expensive edge, right? As you step up to the next level of recursion, now your former root has a parent, and some siblings. For each of the $s$ subtrees connected to the new root, only 1 leaf in each subtree is connected to the new root. For each of these $s$ leaves, you examine the edge from the leaf to the old root, and the edge from the old root to the new root. Delete $s - 1$ of these edges. – Joe Oct 8 '12 at 18:32
• @Joe Sorry for the late response, this has been side-tracked due to other projects. Since I want to minimize the weights of deleted edges, I would keep the edge with the most weight, and delete the rest. I think I understand what you are trying to say. I am convinced that it is correct. Once I get time to think it over, and implement it, I will update accordingly. I would have accepted this had this been an answer (unless proven otherwise). Thanks! – Paresh Oct 19 '12 at 7:00
• I wrote my previous comments as if the tree not the cut was being minimized. – Joe Oct 20 '12 at 7:57

For each node $v$ with height 1 (all its children are leaves): If $v$ has degree $d$ and $d - 1$ children, then the $d-2$ smallest edges are added to the cut. These cuts are required to separate the children of $v$ from each other. Now, all but one child (leaf) of $v$ has been separated from the rest of the tree.

Now consider all the nodes $v$ with height 2. Each of these nodes with degree $d$ has a single parent edge (assuming that $v$ is not the root), and $d - 1$ paths of length at most two connecting it to the leaves in its subtree. We must remove the minimum edge from exactly $d - 2$ of these paths in order to disconnect these leaves from each other. Add these edges to the cut. Now all but one leaf in $v$'s subtree has been separated from the rest of the tree. This leaf is connected to the parent of $v$ and the rest of the tree by a path of length at most $3$. Later on, we will be required to remove at most one edge on this path, so we don't need to track all edges, just the one of minimum weight.

Now consider all nodes $v$ with height $h$. Each of these nodes with $d - 1$ children has exactly $d - 1$ leaves connected in its subtree. Each of these leaves are connected to $v$ by a path of length at most $h$, but have also stored the minimum weight edge on this path. We must cut the smallest $d - 2$ edges in order to separate these leaves from each other. The remaining leaf is connected to the rest of the tree via a path of length at most $h + 1$, and in constant time we store its new minimum weight edge by comparing the old minimum with the weight of the edge (v, v.parent).

By working through the nodes in height order, we maintain the invariants that:

1. In each subtree, we have selected the minimum cut to separate the leaves in that subtree.
2. Each subtree has exactly 1 leaf connected to the root of the subtree.
3. Each subtree maintains the minimum weight edge on the path to that leaf.
4. Each subtree with $k$ leaves has selected exactly $k-1$ edges for its cut.
5. Each subtree is processed in $O(d)$ time, where $d$ is the degree of the root of the subtree.

Correctness and a running time of $O(n)$ follow from these invariants.

• 1) "If v has degree d and d−1 children, then the d−2 smallest edges are added to the cut." -- here you mean $d-1$ leaves and $d-2$ of the edges connecting those, I guess? 2) Paragraph one is redundant? ( 3) For the presentation, note that the tree was not assumed to be rooted.) I don't quite follow your idea, in particular why you find the minimum number. But that might be my tiredness. – Raphael Oct 30 '12 at 21:44
• @Raphael 1) Yes, that is what he means. 2) Paragraph 1 was the base case, paragraph 2 is the generalization. Any node of height more than 1 might have leaves as its descendants, and the path to each leaf will consist of more than one edge. For each such path, we consider the smallest edge in it. This is the edge that needs to be removed to cut off this path with minimum cost. And among these $d-1$ paths, we cut-off the $d-2$ paths with the min edge costs. 3) Any internal (non-leaf) node can be considered to be the root. If no such node exists, the tree is just two nodes with one edge - trivial – Paresh Oct 30 '12 at 22:27
• @Joe From your comments to the question, I had deduced this, and tried to implement it using a single DFS. It did not work out. I must have made a mistake in implementing it, or in the reduction of my original problem to this problem. I will explore both and update. – Paresh Oct 30 '12 at 22:31

Consider this graph with $2k$ nodes:

[source]

From how you describe how you choose the edges to be removed, it is not clear what should happen. If for all pairs $(1,k), (2,k), \ldots, (k-1,k)$ the same node is chosen, you are lucky. But you might also end up removing all the edges with weight $1$, which would leave $k-1$ of the leaves connected. You see that you have to proceed iteratively, choosing the next edge to be removed knowing which are already gone.

Note that the removal of any edge causes a new connected component to arise; thus, removing any $k-1$ edges creates $k$ connected components. So we can just sort edges by weight and remove the $k-1$ smallest -- all we have to take care of is to not remove an edge that creates a new component without any of the original leaves.

To this end, collect in a preprocessing step for each edge all leaves reachable from one and all leaves reachable from the other incident node (without using the considered edge). This takes time $O(m^2)$ and $\Theta(mk)$ extra space, if $m$ is the number of edges (and $n$ the number of nodes), we find reachable leaves by depth-first search and we store the set of reachable leaves per edge in a bit vector.

Now sort the set of edges in time $\Theta(m\log m)$ and perform the greedy algorithm. After removing an edge $(u,v)$, update the bit vectors of all edges in the separated components accordingly: all edges in component $U \ni u$ can no longer reach those leaves reachable from $v$ and vice versa. This update takes time $O(mk)$. The greedy algorithm can not remove $(u,v)$ if $u$ or $v$ can not reach a leaf (without using $(u,v)$).

This algorithm runs in time $O(m^2k)$ (at most $m$ phases of the greedy algorithm each of time $O(mk)$, which dominates everything else). A more precise analysis might make use of the fact that components shrink quickly (on average). Also, the preprocessing step can be sped up by using message passing: the sets of reachable leaves can be propagated among inner edges/nodes.

It is clear that the algorithm separates the $k$ leaves by removing $k-1$ edges. Assume that it did not find an optimal selection of edges. Then, there is an edge $e^*$ separating two sets of leaves $L_1, L_2$ which has smaller weight than one of the edges $e$ on the paths between leaves from $L_1$ and $L_2$ chosen by the algorithm. This is a contradiction to how the algorithm works: it would have considered $e^*$ before $e$ because of its smaller weight, and since it separates a superset of the same leaves, could and would have removed it. Thus, the algorithm is correct.

• Thanks! Yes, this looks correct. Can you please elaborate on how the "message passing" in the preprocessing phase of your algorithm works? I have a feeling that one or two DFS should let us generate the information of the preprocessing phase. But I am not sure how. – Paresh Oct 29 '12 at 12:57
• Also, how did you add the image in your answer? I looked at the source and realized that it is a latex module to help you draw images! How do you include it in your answer? Does simply adding it in between '$' work? – Paresh Oct 29 '12 at 14:12 • @Paresh Message passing can be implemented as recursive algorithm here, because the graph is a tree. If the considered node is a leaf, the edge can reach exactly this node. If not, it can reach all those the edges incident of the considered can reach. Thus, the preprocessing can indeed be done in time$O(mk)$. – Raphael Oct 29 '12 at 14:44 • @Paresh Regarding the image, the source file linked can be compiled locally following these instructions. The resulting image is included just like any other (using the image button in the post editor). – Raphael Oct 29 '12 at 14:46 • The algorithm looks correct, but I am worried about the complexity.$O(m^2k)\$ is a bit too much. I wonder if there could be a faster algorithm. – Paresh Oct 30 '12 at 16:24