You are given a tree with $n$ nodes and $n-1$ edges. Each edge has a weight. There are $m$ queries. In each query you have to answer the following question:

  • How many edges are there in the path between node $a$ and node $b$ with weight less than $k$?

$a$, $b$, and $k$ vary in each query. You're allowed to preprocess the tree to make it easy to answer each query.

Consider the tree with edges

 weight=4/ \weight=5
        2   3

Between nodes 3 and 5, the total number of edges with weight less than 5 is 3, i.e. {1-2, 2-4, 4-5}.

I was trying to solve by brute force method by checking each path.

I want to know if there is a better way of solving this.


You can answer each query in time $O(\log^3 n)$, plus whatever time is needed to answer a lowest-common-ancestor query (a tricky $O(1)$-time solution exists, but you might as well go with the easier $O(\log n)$-time solution, since it doesn't hurt the final time complexity), after doing $O(n\log^3 n)$ preprocessing.

We only need counts from each vertex to the root

Let's call the number of edges of weight less than $k$ on the path from a vertex $v$ to the root $f(v, k)$. We can calculate the desired final answer -- the number of edges of weight less than $k$ on the path from $a$ to $b$ -- using $f(a, k) + f(b, k) - 2f(lca(a, b), k)$, where $lca(u, v)$ is a the lowest common ancestor of $u$ and $v$.

From now on we will focus on answering the simpler question of computing $f(v, k)$.

Heavy-light decomposition

The heavy-light decomposition decomposes the tree into disjoint paths in such a way that at most $O(\log n)$ such "heavy paths" are visited on the path from any vertex to the root. This can be accomplished with a DFS in $O(n)$ time. The idea here is that, since the heavy paths are disjoint, if we have a way to quickly compute the number of edges of weight less than any given $k$ for each one, then we can nearly compute $f(v, k)$ for any $v$ and $k$ just by summing these counts over all heavy paths that we visit on the way from $v$ to the root. And we do have a way to do this counting: Simply by sorting the edges within each heavy path by weight in a preprocessing step, we can later binary-search for any $k$ to count the number of edges of weight less than $k$ in this particular heavy path in $O(\log n)$ time.

But why only "nearly"? Because we have so far ignored the possibility that $v$ might not begin exactly at the bottom of a heavy path -- it could begin part-way up. We could try to "patch things up" by walking down to the bottom of the heavy path containing $v$, subtracting 1 from the count for each edge of weight less than $k$ that we see -- but this won't be worst-case efficient, since an individual heavy path can be as long as $n-1$ edges. We need a way to handle this bottommost heavy path efficiently.

As noted by commenter xskxzr, the issue of beginning part-way up a heavy path is not only confined to the bottommost vertex $v$ in a path to the root -- it can also happen in the middle of such a path, such as in the path from 11 to 0 in the figure on the linked page. Fortunately, the same trick, described next, works for both cases.

Fenwick trees of sorted lists

The subproblem we are now focussed on is: Given the ordered list of edge weights $x_1, \dots, x_m$ encountered while traversing down a particular heavy path, we want to preprocess this list to enable us to be able to later efficiently answer queries of the form "How many of the first $r$ elements are less than $k$?", with both $r$ and $k$ part of the query. One way is to use a Fenwick tree which has as its elements not individual integers, but sorted lists of integers.

A "regular" Fenwick tree with single-integer elements takes $O(n)$ space. Since each $x_i$ belongs to at most $O(\log n)$ buckets, increasing the size of the Fenwick tree's elements from single integers to sorted lists of integers fortunately only increases space usage by a $\log n$ factor. The associative operation in a "regular" Fenwick tree is usually addition; in our Fenwick tree it's "set union". For efficiency, all lists should be populated in a first pass, and then only sorted once, in a second pass. This can be done in $O(n\log^2 n)$ time for a single heavy path, so $O(n\log^3 n)$ overall.

To process an $(r, k)$-query for a particular heavy path, a subquery in that heavy path's Fenwick tree is performed on each of the sequence of buckets found by stripping least-significant 1-bits from $r$, and the counts from these subqueries are summed. That much is standard for Fenwick trees, but the individual subqueries are more complicated: Instead of just reading out the integer value from each bucket, we must binary-search the sorted list in that bucket for $k$; the leftmost position where $k$ appears (or would appear) tells us the number of edges of weight less than $k$ in that bucket.

An example might help: Suppose $r = 43$. In binary, 43 is 101011 (32 + 8 + 2 + 1). To answer a query for $r = 43$ (and some $k$ value), we would binary-search for $k$ in the following 4 buckets: 101011, 101010, 101000, 100000. Each bucket is found by stripping the least-significant 1-bit from the previous bucket, so there cannot be more than $\log n$ subqueries. Each subquery is a binary search, so cannot take longer than $O(\log n)$ time. Finally, there are at most $O(\log n)$ heavy paths encountered on the path from any vertex to the root, so the worst-case time to compute $f(v, k)$ is $O(\log^3 n)$.

  • 1
    $\begingroup$ Not only $v$, but also some intermediate node might not lie in the bottom of a heavy path. $\endgroup$
    – xskxzr
    Feb 27 '19 at 6:56
  • $\begingroup$ @xskxzr: I don't quite follow; by "intermediate node", do you mean the topmost node in the path between the two given vertices $a$ and $b$? We have already reduced this original problem to (3 instances of) the special case of finding the number of edges with weight below the given threshold on the path from a given vertex $v$ to the root. $\endgroup$ Mar 2 '19 at 5:42
  • 1
    $\begingroup$ For example, in the example graph in the link you provided, consider the path from 11 to 0, the intermediate node 2 is not in the bottom of a heavy path. Of course, your algorithm is correct. $\endgroup$
    – xskxzr
    Mar 2 '19 at 5:55
  • $\begingroup$ @xskxzr: Thanks, I understand now -- you're right, and that's something I overlooked. Fortunately, as you noted, the same special handling applied to $v$ can be applied to these intermediate nodes. I'll edit in a moment. $\endgroup$ Mar 5 '19 at 10:04
  • $\begingroup$ "Given the ordered list of edge weights ..." Shouldn't it be unordered? $\endgroup$
    – xskxzr
    Mar 5 '19 at 10:26


This is an alternative to j_random_hacker's solution, or more precisely, an alternative to his/her solution to the subproblem:

Given the ordered list of edge weights $x_1,\ldots,x_m$ encountered while traversing down a particular heavy path, we want to preprocess this list to enable us to be able to later efficiently answer queries of the form "How many of the first $r$ elements are less than $k$?"

j_random_hacker uses the data structure of Fenwick tree of sorted lists, which results in a solution with

  • overall $O(n\log^3 n)$ preprocessing time for all heavy paths,
  • overall $O(n\log n)$ space for all heavy paths, and
  • $O(\log^2 n)$ query time for the subproblem, hence $O(\log^3 n)$ query time for the primary problem.

In contrast, my alternative gives a solution with

  • overall $O(n\sqrt{n})$ preprocessing time for all heavy paths,
  • overall $O(n\sqrt{n})$ space for all heavy paths, and
  • $O(1)$ query time for the subproblem, hence $O(\log n)$ query time for the primary problem.

How to preprocess each heavy path?

In the very first, before we turn to those heavy paths, we sort all edges according to their weights from small to large. Say the result is $e_1,\ldots,e_{n-1}$ with weights $w_1,\ldots,w_{n-1}$. We label edge $e_i$ by $i$ for future use. Denote $I_1=(-\infty,w_1],I_2=(w_1,w_2],\ldots,I_n=(w_{n-1},+\infty)$.

Now let's focus on the subproblem mentioned above. We divide $x_1,\ldots,x_m$ into blocks of length $t=\lceil\sqrt{n}\rceil$: $\left[x_1,\ldots,x_t\right],\left[x_{t+1},\ldots,x_{2t}\right],\ldots$ We build two tables $T_1, T_2$ where $T_1(p,q,r)$ represents how many of the first $r$ elements in the $q$th block are less than $k$ if $k\in I_p$, and $T_2(p,q)$ represents how many of the elements in the first $q$ blocks are less than $k$ if $k\in I_p$.

Now if we have these two tables, we can answer an query for each heavy path in $O(1)$ time. For example, for a query $(r,k)$, if $k\in I_5$ and $r=3t+2$, then the answer is $T_1(5,4,2)+T_2(5,3)$.

Note it takes $O(\log n)$ time to find $p$ such that $k\in I_p$. Since we need to do this search only once for each query of the primary problem, it does not increase the query time.

How to build these tables?

Note $T_2(p,q)=\sum_{i=1}^q T_1(p,i,t)$, if we already have $T_1$, we can build $T_2$ in $O(nm/t)$ time and it takes $O(nm/t)$ space. Next we focus on how to build $T_1$.

Suppose the edge weights in the $q$th block are $w_{i_1},\ldots,w_{i_t}$ where $i_1<\cdots<i_t$ (recall that these edges are already labeled by $i_1,\ldots,i_t$). Then we have

\begin{align} T_1(1,q,*)=T_1(2,q,*)=&\cdots=T_1(i_1,q,*),\\ T_1(i_1+1,q,*)=T_1(i_1+2,q,*)=&\cdots=T_1(i_2,q,*),\\ &\cdots\\ T_1(i_t+1,q,*)=T_1(i_t+2,q,*)=&\cdots=T_1(n,q,*), \end{align}

where $T_1(p,q,*)$ represents the array $[T_1(p,q,1),\ldots,T_1(p,q,t)]$, and two arrays are equal if they are element-wise equal. This means we only need to compute and store $t$ arrays for $T_1(1, q, *),\ldots,T_1(n, q, *)$. So we can build $T_1$ in $O(t^2\cdot m/t)=O(mt)$ time, and it also takes $O(mt)$ space.

Note $t=\lceil\sqrt{n}\rceil$, the overall preprocessing time for all heavy paths are

$$ \sum_m O(m\sqrt{n})=O(n\sqrt{n}). $$

So is the overall space used.


Judging by the comments so far, your best bet is probably to build a dictionary with each pair of nodes as keys, and any info about their path that you need.

You can build this in a systematic way by traversing the tree and reusing the partial paths as you go on. This way you can reuse partial results to improve on time complexity.

Being more specific: use Floyd-Warshall algorithm to systematically explore all shortest paths between each pair of nodes. Since this is a tree, there will only be one such path for each pair.
While computing the partial solutions, also keep an extra matrix that keeps track of how many edges are less than 5.
Any time you include a new path the number of edges with weights less than 5 is the sum of the number of such edges in the partial solutions. If the current edge being added is less than 5, add 1 to this calculated value.

  • $\begingroup$ I've edited the answer converting your problem to an instance of the all-pairs shortest-paths. For the APSP problem algorithms there are a lot of references out there! :) $\endgroup$ Dec 28 '18 at 17:41
  • $\begingroup$ Note that “5” is an input, not a fixed number. $\endgroup$ Dec 28 '18 at 18:38
  • $\begingroup$ hello sir , i think u didn't get the question since there are n nodes and n-1 edges therefore the given graph is a tree therefore path between every two node is unique and is shortest then why should i use floyd warshall algorithm which has certainly has a large time complexity i.e o (nnn) $\endgroup$ Dec 29 '18 at 11:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.