**Here is a playlist of all hybrid tutorials I've done.**

# "Intro"

*Timestamp: 00:00*

Hi! Definitely not inspired by this comment, I've decided to try something that seems relatively novel — combining a blog and video tutorial into one, in a "hybrid" fashion. Both should be usable independently, but they will have the same "flow" and structure so you can reference both for concepts that are harder to grasp, and the two will supplement each other. The goal of these is to be **complete** — beneficial for both video and blog lovers, as well as full of enough information that anyone without much of an idea of what the concept is should be able to fully understand. There will be code as well, however, I very highly recommend not looking at it, but rather working out the implementation for yourself.

The concept of this tutorial is heavy-light decomposition. This one isn't 100% fair, as I'm basically copy-pasting one of my old tutorials that can be found here and adding a video component, but it's never been on Codeforces before.

The full video can be found here.

# The motivation (path queries)

*Timestamp: 01:20*

Heavy-light decomposition is kind of a scary name, but it's really cool for what it does. Take some classic segment tree problem with two types of queries:

- Compute the XOR of values on the subarray from $$$l$$$ to $$$r$$$
- Update the value at position $$$i$$$ to $$$x$$$

Now, anyone familiar with a segment tree will recognize that it can be easily applied to solve this. But what if we moved the problem from queries on an array to queries on a tree? Let's reformulate the queries to be:

- Compute the XOR of values on the path from $$$u$$$ to $$$v$$$
- Update the value at vertex $$$i$$$ to $$$x$$$

which is exactly this problem from USACO, and the problem I'll be using as an example for the rest of this tutorial. To note, its editorial also provides a nice description of HLD.

# Prerequisites

*Timestamp: 05:05*

This is a relatively advanced topic, and this tutorial is not meant for beginners. Before reading this blog, you should know:

- Basic tree algorithms (such as dfs)
- Segment trees (at the very least, what their purpose is), with lazy propagation
- Binary lifting (useful for my implementation, other implementations may not use it)
- Lowest common ancestor (LCA)

# The concept

*Timestamp: 07:50*

So... what? How do we turn a tree into something we can use a segment tree on? Well, we can't exactly. What we *can* do is turn a tree into a bunch of paths, each of which we can build a segment tree on. That's an incredibly vague description, so let me describe what I mean.

Consider this poorly drawn tree (rooted arbitrarily):

What we're going to do is, for each vertex that isn't a leaf, find the edge that leads to the child with the largest subtree (breaking ties arbitrarily). I'll color these edges green:

(blame MS Paint for the colors looking weird). We'll call these special edges "heavy" edges (because they lead to the "heaviest" child) and the rest of them "light" edges. This turns out to be super nice to work with because the heavy edges always form "vertical chains" — paths that go from some vertex to an ancestor of it. Brushing aside the implementation for a second, let's assume we've built a segment tree on each of these vertical chains. Now the key idea about this setup is that **any path on the tree will pass through at most $$$O(\log{n})$$$ light edges**. By extension, each path also passes through at most $$$O(\log{n})$$$ vertical chains.

We should definitely prove that. For notation, from now on I will call paths of heavy edges "heavy chains". One useful idea for the proof of this claim is that you can break any path $$$u$$$ -> $$$v$$$ on a tree into two (possibly non-existent) components: the path from $$$u$$$ up to $$$lca(u, v)$$$ and the path from $$$v$$$ up to $$$lca(u, v)$$$, where $$$lca(u, v)$$$ is the lowest common ancestor of $$$u$$$ and $$$v$$$. Because $$$lca(u, v)$$$ is an ancestor of both $$$u$$$ and $$$v$$$, both of these separate paths will also be vertical chains themselves. So now let's prove that both of these vertical chains only pass through $$$O(\log{n})$$$ light edges.

**Proof**

Cool. Now, since each path goes through at most $$$O(\log{n})$$$ light edges and heavy chains, we just need to quickly evaluate all of these. Light edges are single edges, so we can get their vertices' values quickly. As I said before, if we build a segment tree on each heavy chain, then we can evaluate the value of the intersection between our path and each heavy chain in $$$O(\log{n})$$$. Thus, the total complexity is $$$O(\log^2{n})$$$ per query.

# My implementation

*Timestamp: see below*

This is the fun part. I'll break this up into sections. The tree is always rooted arbitrarily (at 0).

**Finding the heavy edges**

*Timestamp: 20:20*

We compute subtree sizes with a simple DP: $$$sz[v] = 1 + \sum\limits_{u \in child(v)} sz[u]$$$ where $$$child(v)$$$ is the set of all children of $$$v$$$. Then we just have to take the child $$$u$$$ with the maximum value of $$$sz[u]$$$ of all children, for each vertex. In this dfs, I also set the depths of each vertex relative to the root, which is used later for evaluating queries.

**Code**

**Building a segment tree on each heavy chain**

*Timestamp: 24:10*

For convenience (so I or someone else can stick a segment tree template in), I choose to use a single segment tree for everything and label the nodes so heavy chains end up together in the labeling. We do a dfs and label nodes in the order that they're visited. The simplest way to achieve labeling like what we need is to visit big children before everything else, then the label of each big child will be exactly 1 greater than that of its parent. Here's what this would look like on the tree above:

As you can see, for this tree (and all other trees), the labels on heavy chains form consecutive sequences of numbers, so they can be easily queried with the segment tree.

**Code**

**Finding heavy chains**

*Timestamp: 25:48*

Now our goal is to compute, for each vertex, the vertex at the top of the heavy chain it belongs to (which could be itself). This can be done in a DP-like fashion going down the tree. We set $$$chain[i]$$$ to:

- $$$chain[par[i]]$$$, if $$$i$$$ is the heavy child of its parent
- $$$i$$$, otherwise.

because if $$$i$$$ is a light child, then there's clearly no heavy chain above it, and otherwise, it just continues the heavy chain from its parent.

**Code**

**Computing queries**

*Timestamp: 28:37*

*Timestamp for vertical chains, specifically: 29:57*

The best part, of course. We will, once again, leverage the fact that the path between any $$$u$$$ and $$$v$$$ is the same as the path from $$$u$$$ to $$$lca(u, v)$$$ combined with the path from $$$v$$$ to $$$lca(u, v)$$$. The algorithm is essentially: go up from $$$u$$$ (inclusive) to $$$lca(u, v)$$$ (exclusive) by efficiently computing the values from light vertices and heavy chains, then do the same for $$$v$$$. The convenience of setting $$$chain[u] = u$$$ is that we can do the same thing regardless of the edge above a vertex being light or heavy. Use the segment tree to compute the value of the path from $$$u$$$ to $$$chain[u]$$$ then set $$$u := par[chain[u]]$$$ until you get to computing the vertex that's a parent of $$$u$$$ and directly below $$$lca(u, v)$$$ (because we need to exclude $$$lca(u, v)$$$).

Wait... how do we get the vertex directly below $$$lca(u, v)$$$ (which has to be a parent of $$$u$$$ or $$$v$$$)? The smart way to do this is to use the labels we've already constructed because as we already know, they're consecutive. I am not smart, so I used binary lifting to get the $$$(dep[u] - dep[lca(u, v)] - 1)$$$-th ancestor of $$$u$$$. Ultimately these have the same effect, and the binary lifting only adds another $$$O(\log{n})$$$ to the query because we only need it once.

Now that we have that vertex, we need to make sure we don't accidentally include the LCA or vertices above it in the query. We can do that by checking each time, if $$$depth[chain[v]] \leq depth[lca(u, v)]$$$. Once this is true, we set our top node (which would usually be $$$chain[v]$$$) to the vertex directly below $$$lca(u, v)$$$, which automatically means we stop next iteration since we reach $$$lca(u, v)$$$.

**Code (computing a vertical chain)**

To get the answer for the full query, we combine the answers for the two vertical chains and, additionally, include the LCA.

**Code (computing a full query)**

**Applying updates**

This is exactly the same as computing a query, but replace `seg_query_header`

with `seg_update_header`

.

# The full implementation

The implementation specifically applied to the USACO problem (submitting that exact code will give AC) is here. An implementation with just the HLD is here. Some more details follow.

Note: This implementation is by no means optimized in terms of constants. Of course, the asymptotic complexity is the expected $$$O(\log^2{n})$$$ per query, but there are many things that can be improved. I'll leave these optimizations as a task for you.

**How to initialize it**

- Decide on appropriate values for the template header (should be a value larger than what'll ever be necessary for array sizes)
- Read in the value of $$$n$$$, and call
`init_arrays(n)`

- Read in the edges, and call
`add_edge`

($$$0$$$-indexed) for each one - Call
`init_tree`

and pass it an array with the initial values (replace with vector pointer if you want).**The root is not important, and can be arbitrary.**

And you're done! Now you can call `query`

and `update`

with the endpoints of paths (also $$$0$$$-indexed).

**The segment tree**

Most of the stuff will work on its own, you won't have to worry about it. I've marked out a section that should be changed depending on the problem — such as segment tree combination (in the case of the USACO problem, xor), sentinel values, and how to handle lazy propagation. Of course, feel free to replace this with your own segment tree.

- The
`seg_combine`

function is the function you want to use to combine values on the segment tree. As with a normal segment tree, it has to be a monoid (associative, etc.) - The
`seg_lazy_apply`

function should be what you do to "enqueue" a lazy update (without actually applying it yet, which is kinda counterintuitive for the function name). In this example, because our operation is a set operation, we would set`lazy_val`

to`new_val`

and just return`new_val`

. If we wanted range add, we would return`lazy_val + new_val`

. - The
`seg_lazy_func`

should be what you do to actually apply the updates. In this example, a simple set operation is enough because we have point, not range, updates. If we were doing, for example, range set with a sum segment tree, the appropriate value would be`lazy_val * (r - l + 1)`

because each element on the range $$$[l, r]$$$ contributes exactly`lazy_val`

to the sum. `seg_sentinel`

is the value used for queries that return the range. It should be a value that should have**no effect**on the return value from`seg_combine`

(usually called the "identity"). For example, for a max segment tree, this would be a very large (by magnitude) negative number, and with sum or xor it should be`0`

.`seg_lazy_sentinel`

is a flag for lazy propagation, saying that no operation should be done. This is especially important for range set queries, as if the sentinel is something like`0`

, the segment tree would always think everything should be set to`0`

. It should be set to a value that can never possibly be an actual value in the lazy array.`seg_init_val`

should almost always be`0`

, it's just a starting value before initializing with the actual values in`dfs_labels`

.

**Templates and seg_t**

The value of `size`

in the template should be at least as large as the arrays will ever have to get, and `lg`

should always be greater than $$$\log_2{(size)}$$$ for binary lifting purposes. `seg_t`

is used all over, and is essentially the data type that should be used for queries, updates, and general segment tree operations. Usually, it can be `long long`

, but if speed (or memory) becomes an issue you can make it `int`

or something else as well (including custom data types, like matrices and stuff).

**Edge (instead of vertex) queries/updates**

*Timestamp: 37:13*

We can easily modify this HLD to work with edge queries instead of vertex ones. What we can do is, we can leverage the property that each vertex except the root has an edge going up to its parent. That means we can use the values in each vertex to represent the edges going from them to their parents without modifying most of the code. This only requires two real slight changes:

- Modify the
`query`

and`update`

functions to exclude the LCAs, since we don't want to include the edge from the LCA to its parent - Carefully set up initialization so that the value at each position in the array represents the edge going up from the vertex at that position (the value at the root $$$0$$$ can be anything, as it will always be ignored)

**Verification/practice problems**

First tested on the USACO problem, linked above.

I also used my HLD to solve this problem (solution link), which you can try as a challenge if you want. To my surprise, it ran rather quickly!

You can use the above two problems as well as:

- Path Queries (CSES)
- QTREE (SPOJ): allows you to test modifications for edges
- GRASSPLA (SPOJ; original source is USACO but the judge doesn't work for that problem)
- GSS7 (SPOJ)
- QRYLAND (CodeChef)
- MONOPLOY (CodeChef)
- QUERY (CodeChef)
- BLWHTREE (CodeChef)
- Milk Visits (USACO)
- Max Flow (USACO)
- Exercise Route (USACO)

Note that these are not ordered by difficulty (if they were, it would be a bit of a spoiler) and that HLD may not necessarily be the best/simplest solution to these problems, but it's possible to solve these with it. Also, feel free to suggest more practice problems, as I don't really know where to find them.

edit: Continuing with the relatively advanced graph algorithms (as well as the "decomposition" techniques), the next topic will be centroid decomposition. I'll publish it sometime this or next week.

Bro this was actually some quality content I sincerely appreciate the hard work done. I finally got to understand HLD only after watching your video, great job!

Absolute gold.....I wish there was a way where people like you, errichto, secondthread etc...could colab with the edu section of condeforces....would essentially make everything much more centralised.

Thanks a lot for all your tutorials.

EDIT: what follows is completely wrong, the height is not necessarily O(log(n)). A binary tree where every left son is leaf has an height O(n)

I have a slightly different method and I wondered if it's common and whether it has a name. I consider as "heavy" the edges from a parent node to a son which itself as only one son. Heavy chain are created too. When you compress a heavy chain as one edge you get a tree where each node is either a leaf or got 2 or more sons. Therefore the height is O(log(n)). So you can apply the same things with same complexity, except that all heavy chains (except the starting ones with u and v) are complete so queries on them are computed in O(1) : the whole query is computed in O(log(n)). I didn't implement it so I may be completely wrong. Let me know if it's unclear or if I made any mistakes. Best regards,

In your implementation I was trying to find where the elements of lca_lift were being initialised to -1 (actually trying to find a loop for it). I was amazed when I understood how it was being done smartly without having to initialise every element to -1 from the start. P.S. very nice explanation through this hybrid tutorial :).

https://acm.timus.ru/problem.aspx?space=1&num=1553