# Finding median weights in all paths of an AVL tree with weighted nodes

I need some help with proving the complexity of the following problem (I'm new here, so please excuse my "newbie-ness")

Given: an AVL tree with keys: $1,2,..,n$, such that each node $i$ in the tree has the parameter $w_i$ (which is an integer)

The task: To make an algorithm which returns an array of size $n$ such that in the cell $i$ there is the median value of all the $w_i$ of all the nodes from the root of the AVL to the node $i$.

Running time: $O(n\log\log n)$

My solution:

• For each node among the $n$ nodes in the AVL, I build a separate AVL containing the $O(\log n)$ nodes(because the depth of each node is $O(\log n)$) on the path to it from the root.
• Each insertion to an AVL takes $O(\log n)$ ($n$ = number of nodes), but number of nodes here is $O(\log n)$, therefore, each insertion takes $O(\log\log n)$, and since there are $n$ nodes, then building the $n$ AVLs takes $O(n\log\log n)$
• Finding the median for each AVL takes $O(\log n)$ (using Select) since there are $O(\log n)$ values in the tree. Then, we put in the output array at the cell $i$ the median value which we just found.

Thus, building each AVL and finding the median, is done in $O(\log(n) + \log\log n) = O(\log n)$.

My Question: I'm uncertain about step 2 in the described algorithm above: does building the $n$ AVL trees takes actually $O(n\log\log n)$ running time?

• $\log n + \log\log n$ is $O(\log n)$ but not $O(\log\log n)$ so, without any further checking, your analysis seems to be flawed. Jun 17 '15 at 10:16
• right..I just paid attention to that. But how could I get past that? Jun 17 '15 at 11:33
• First you should correct the $O(\log(n) + \log\log n) = O(\log\log n)$ into ...$=O(\log n)$, as suggested by @D since it is painful to read. Then, your step 2 does seem incorrect. You have $n$ small AVL trees, each insertion is bound by $\log\log n$ cost since they have each at most $log n$ size, but you have $\log n$ nodes to insert in each, which makes a total cost $O(n\log n\log\log n)$ ... BTW, what is your definition of median value. Jun 17 '15 at 14:13
• a median is: ceiling((n+1)/2) element in size Jun 17 '15 at 15:02

Your main AVL tree has $n$ nodes. On each you build a small AVL tree with a nomber of nodes bounded by $\log n$, so that the insertion of each node on the small AVL tree has a cost $\log\log n$ steps. Overall that makes $O(n\log n\log\log n)$ steps, not $O(n\log\log n)$ as stated in your step 2.

• The way I thought about it is as follows: since each node belongs to a small AVL, and it took $O(\log\log n)$ to insert it there(to the small AVL), thus, since we got $n$ nodes, it should take $O(n\log\log n)$. where is my mistake here? Jun 17 '15 at 16:02
• What I understand from your construction is that each node is inserted in several small AVL trees, Jun 17 '15 at 16:07
• that's correct. Jun 17 '15 at 16:11
• @ThunderWiring Is it correct that you build $n$ small AVL. Is it correct that each contain on the order of $\log n$ nodes (You say so yourself). So that makes $0(n\log n)$ nodes to insert, each in $O(\log\log n)$ steps. Hence the answer I gave you. I do not see what is still a problem. ... but though I did not work it out in detail, I would not try to solve your problem the way you do it. Jun 17 '15 at 19:27

Maintaining one AVL tree for the whole process is enough, and more efficient.

• Setup another AVL tree T, initially empty.
• Perform DFS on original AVL tree
• Insert node to T on entering the node,
• Delete node from T on leaving the node.
• Find and record the median after inserting each node.

In the DFS process, T actively maintains all nodes along the path from root to current node, similar to the stack (explicitly or implicitly) maintained in DFS.

The size of T is $O(\log n)$ during the whole process, so each insertion/deletion/query cost $O(\log \log n)$. And each node in original AVL is inserted/deleted/queried exactly once. So the overall complexity is $O(n \log \log n)$.

The maintained AVL tree T can be replaced by other running median data structures. A common data structure min heap + max heap, maintaining the 2 halves partitioned by current median, which can report current median in $O(1)$. (insertion/deletion is still $O(\log n)$)

• Welcome to Computer Science Stack Exchange. Please read cs.stackexchange.com/tour, if you have not yet done so. Read questions carefully before answering. This question is not about a better algorithm but about the asymptotic cost of the solution presented in the question. Some users might downvote for that. But you do give the answer the original poster was looking for, though he did not ask it. Aug 11 '15 at 10:14