I've heard of a new balanced BST data structure called a zip tree. What is the zip tree? How does it work?
At a high level, a zip tree is a
randomized balanced binary search tree,
that is a way of encoding a skiplist as a BST, and
that uses a pair of operations called zipping and unzipping rather than tree rotations.
The first bullet point - that zip trees are randomized, balanced BSTs - gives a feel for what a zip tree achieves at a high level. It's a type of balanced binary search tree that, like treaps and unlike red/black trees, uses randomization to balance the tree. In that sense, a zip tree isn't guaranteed to be a balanced tree, but rather has a very high probability of being balanced.
The second bullet point - that zip trees are encodings of skiplists - shows where zip trees come from and why, intuitively, they're balanced. You can think of a zip tree as a way of taking the randomized skiplist data structure, which supports all major operations in expected time O(log n), and representing it as a binary search tree. This provides the intuition for where zip trees come from and why we'd expect them to be so fast.
The third bullet point - zip trees use zipping and unzipping rather than tree rotations - accounts for the name of the zip tree and what it feels like to code one up. Zip trees differ from other types of balanced trees (say, red/black trees or AVL trees) in that nodes are moved around the tree not through rotations, but through a pair of operations that convert a larger chain of nodes into two smaller chains or vice-versa.
The rest of this answer dives deeper into where zip trees come from, how they work, and how they're structured.
Review: Skip Lists
To understand where zip trees come from, let's begin with a review of another data structure, the skiplist. A skiplist is a data structure that, like a binary search tree, stores a collection of elements in sorted order. Skiplists, however, aren't tree structures. Rather, a skiplist works by storing elements in sorted order through several layers of linked lists. A sample skiplist is shown here:
As you can see, the elements are represented in sorted order. Each element has an associated height, and is part of a number of linked lists equal to its height. All of the elements of the skiplist participate in the bottom layer. Ideally, roughly half of the nodes will be in the layer above that, roughly a quarter of the nodes will be in the layer above that, roughly an eighth of the nodes will be in the layer above that, etc. (More on how this works later on.)
To do a lookup in a skiplist, we begin in the topmost layer. We walk forward in the skiplist until either (1) we find the element we're looking for, (2) we find an element bigger than the one we're looking for, or (3) we hit the end of the list. In the first case, we uncork the champagne and celebrate because we discovered the item we were searching for and there's nothing more to do. In the second case or third cases, we've "overshot" the element that we're looking for. But that's nothing to worry about - in fact, that's helpful because it means that what we're looking for must be between the node we hit that "overshot" and the node that comes before it. So we'll go to the previous node, drop down one layer, and pick up our search from there.
For example, here's how we'd do a search for 47:
Here, the blue edges indicate the links followed where we moved forward, and the red edges indicate where we overshot and decided to descend down a layer.
A powerful intuition for how skiplists work - which we'll need later on as we transition to zip trees - is that the topmost layer of the skiplist partitions the remaining elements of the skiplists into different ranges. You can see this here:
Intuitively, a skiplist search will be "fast" if we're able to skip looking at most of the elements. Imagine, for example, that the second-to-last layer of the skiplist only stores every other element of the skiplist. In that case, traversing the second-to-last layer is twice as fast as traversing the bottom layer, so we'd expect a lookup starting in the second-to-last layer to take half as much time as a lookup starting in the bottom layer. Similarly, imagine that the layer above that one only stores every other element from the layer below it. Then searching in that layer will take roughly half as much time as searching the layer below it. More generally, if each layer only stores roughly half the elements of the layer below it, then we could skip past huge amounts of the elements in the skiplist during a search, giving us good performance.
The skiplist accomplishes this by using the following rule: whenever we insert an element into the skiplist, we flip a coin until we get heads. We then set the height of the newly-inserted node to be the number of coins that we ended up tossing. This means it has a 50% chance to stay in its current layer and a 50% chance to move to the layer above it, which means, in aggregate, that roughly half the nodes will only be in the bottom layer, roughly half of what's left will be one layer above that, roughly half of what's left will be one layer above that, etc.
(For those of you with a math background, you could also say that the height of each node in the skiplist is a Geom(1/2) random variable.)
Here's an example of inserting 42 into the skiplist shown above, using a height of 1:
Deletion from a skiplist is also a fairly simple operation: we simply splice it out of whatever linked lists it happens to be in. That means that if we were to delete the 42 we just inserted from the above list, we'd end up with the same skiplist that we started with.
It can be shown that the expected cost of an insertion, deletion, or lookup in a skiplist is O(log n), based on the fact that the number of items in each list is roughly half the number of items in the one below it. (That means we'd expect to see O(log n) layers, and only take a constant number of steps in each layer.)
From Skiplists to Zip Trees
Now that we've reviewed skiplists, let's talk about where the zip tree comes from.
Let's imagine that you're looking at the skiplist data structure. You really like the expected O(log n) performance of each operation, and you like how conceptually simple it is. There's just one problem - you really don't like linked lists, and the idea of building something with layers upon layers of linked lists doesn't excite you. On the other hand, you really love binary search trees. They've got a really simple structure - each node has just two pointers leaving it, and there's a simple rule about where everything gets placed. This question then naturally arises: could you get all the benefits of a skiplist, except in BST form?
It turns out that there's a really nice way to do this. Let's imagine that you have the skiplist shown here:
Now, imagine you perform a lookup in this skiplist. How would that search work? Well, you'd always begin by scanning across the top layer of the skiplist, moving forward until you found a key that was bigger than the one you were looking for, or until you hit the end of the list and found that there were no more nodes at the top level. From there, you'd then "descend" one level into a sub-skiplist containing only the keys between the last node you visited and the one that overshot.
It's possible to model this exact same search as a BST traversal. Specifically, here's how we might represent the top layer of that skiplist as a BST:
Notice that all these nodes chain to the right, with the idea being that "scanning forward in the skiplist" corresponds to "visiting larger and larger keys." In a BST, moving from one node to a larger node corresponds to moving right, hence the chain of nodes to the right.
Now, each node in a BST can have up to two children, and in the picture shown above each node has either zero children or one child. If we fill in the missing children by marking what ranges they correspond to, we get this.
And hey, wait a minute! It sure looks like the BST is partitioning the space of keys the same way that the skiplist is. That's promising, since it suggests that we're on to something here. Plus, it gives us a way to fill in the rest of the tree: we can recursively convert the subranges of the skiplist into their own BSTs and glue the whole thing together. If we do that, we get this tree encoding the skiplist:
We now have a way of representing a skiplist as a binary search tree. Very cool!
Now, could we go the other way around? That is, could we go from a BST to a skiplist? In general, there's no one unique way to do this. After all, when we converted the skiplist to a BST, we did lose some information. Specifically, each node in the skiplist has an associated height, and while each node in our BST has a height as well it's not closely connected to the skiplist node heights. To address this, let's tag each BST node with the height of the skiplist node that it came from. This is shown here:
Now, some nice patterns emerge. For starters, notice that each node's associated number is bigger than its left child's number. That makes sense, since each step to the left corresponds to descending into a subrange of the skiplist, where nodes will have lower heights. Similarly, each node's associated number is greater than or equal to the number of its right child. And that again makes sense - moving to the right either means
continuing forward at the same level that we were already on, in which case the height remains the same, or
hitting the end of a range and descending into a subrange, in which case the height decreases.
Can we say more about the shape of the tree? Sure we can! For example, in a skiplist, each node's height is picked by flipping coins until we get heads, then counting how many total coins we flipped. (Or, as before, it's geometrically distributed with probability 1/2). So if we were to imagine building a BST that corresponded to a skiplist, we'd want the numbers assigned to the nodes to work out the same way.
Putting these three rules together, we get the following, which defines the shape of our tree, the zip tree!
A zip tree is a binary search tree where
Each node has an associated number called its rank. Ranks are assigned randomly to each node by flipping coins until heads is flipped, then counting how many total coins were tossed.
Each node's rank is strictly greater than its left child's rank.
Each node's rank is greater than or equal to its right child's rank.
It's amazing how something like a skiplist can be represented as a BST by writing out such simple rules!
Inserting Elements: Unzipping
Let's suppose you have a zip tree. How would you insert a new element into it?
We could in principle answer this question by looking purely at the rules given above, but I think it's a lot easier to figure this out by remembering that zip trees are skiplists in disguise. For example, here's the above zip tree, with its associated skiplist:
Now, suppose we want to insert 18 into this zip tree. To see how this might play out, imagine that we decide to give 18 a rank of 2. Rather than looking at the zip tree, let's look at what would happen if we did the insertion into the skiplist. That would give rise to this skiplist:
If we were to take this skiplist and encode it as a zip tree, we'd get the following result:
What's interesting about this is that we can see what the tree needs to look like after the insertion, even if we don't know how to perform the insertion. We can then try to figure out what the insertion logic needs to look like by reverse-engineering it from these "before" and "after" pictures.
Let's think about what change this insertion made to our zip tree. To begin with, let's think back to our intuition for how we encode skiplists as zip trees. Specifically, chains of nodes at the same level in a skiplist with no intervening "higher" elements map to chains of nodes in the zip tree that lean to the right. Inserting an element into the skiplist corresponds to adding some new element into one of the levels, which has the effect of (1) adding in something new into some level of the skiplist, and (2) taking chains of elements in the skiplist that previously were adjacent at some level, then breaking those connections.
For example, when we inserted 18 into the skiplist shown here, we added something new into the blue chain highlighted here, and we broke all of the red chains shown here:
What is that going to translate into in our zip tree? Well, we can highlight the blue link where our item was inserted here, as well as the red links that were cut:
Let's see if we can work out what's going on here. The blue link here is, fortunately, pretty easy to find. Imagine we do a regular BST insertion to add 18 into our tree. As we're doing so, we'll pause when we reach this point:
Notice that we've hit a key with the same rank as us. That means that, if we were to keep moving to the right, we'd trace out this region of the skiplist:
To find the blue edge - the place where we go - we just need to walk down through this chain of nodes until we find one bigger than us. The blue edge - our insertion point - is then given by the edge between that node and the one above it.
We can identify this location in a different way: we've found the blue edge - our insertion point - when we've reached a point where the node to insert (1) has a bigger rank than the node to the left, (2) has a rank that's greater than or equal to the node on the right, and (3) if the node to the right has the same rank, our new item to insert is less than the item to the right. The first two rules ensure that we're inserting into the right level of the skiplist, and the last rule ensures that we insert into the right place in that level of the skiplist.
Now, where are our red edges? Intuitively, these are the edges that were "cut" because 18 has been added into the skiplist. Those would be items that previously were between the two nodes on opposite ends of the blue edge, but which node need to get partitioned into the new ranges defined by the split version of that blue edge.
Fortunately, those edges appear in really nice places. Here's where they map to:
(In this picture, I've placed the new node 18 in the middle of the blue edge that we identified in the skiplist. This causes the result not to remain a BST, but we'll fix that in a minute.)
Notice that these are the exact same edges that we'd encounter if we were to finish doing our regular BST insertion - it's the path traced out by looking for 18! And something really nice happens here. Notice that
each time we move to the right, the node, when cut, goes to the right of 18, and
each time we move to the left, the node, when cut, goes to the left of 18.
In other words, once we find the blue edge where we get inserted, we keep walking as though we were doing our insertion as usual, keeping track of the nodes where we went left and the nodes where we went right. We can then chain together all the nodes where we went left and chain together all the nodes where we went right, gluing the results together under our new node. That's shown here:
This operation is called unzipping, and it's where we get the name "zip tree" from. The name kinda make sense - we're taking two interleaved structures (the left and right chains) and splitting them apart into two simpler linear chains.
To summarize:
Inserting x into a zip tree works as follows:
Assign a random rank to x by flipping coins and counting how many flips were needed to get heads.
Do a search for x. Stop the search once you reach a node where
the node's left child has a lower rank than x,
the node's right child has a rank less than or equal to x, and
the node's right child, if it has the same rank as x, has a larger key than x.
Perform a unzip. Specifically:
Continue the search for x as before, recording when we move left and when we move right.
Chain all the nodes together where we went left by making each the left child of the previously-visited left-moving node.
Chain all the nodes together where we went right by making each the right child of the previously-visited right-moving node.
Make those two chains the children of the node x.
You might notice that this "unzipping" procedure is equivalent to what you'd get if you performed a different operation. You could achieve the same result by inserting x as usual, then using tree rotations to pull x higher and higher in the tree until it came to rest in the right place. This is a perfectly valid alternative strategy for doing insertions, though it's a bit slower because two passes over the tree are required (a top-down pass to insert at a leaf, then a bottom-up pass to do the rotations).
Removing Elements: Zipping
Now that we've seen how to insert elements, how do we remove them?
Let's begin with a helpful observation: if we insert an item into a zip tree and then remove it, we should end up with the exact same tree that we started with. To see why this is, we can point back to a skiplist. If you add and then remove something from a skiplist, then you end up with the same skiplist that you would have had before. So that means that the zip tree needs to end up looking identical to how it started after we add and then remove an element.
To see how to do this, we'd need to perform two steps:
Undo the unzip operation, converting the two chains of nodes formed back into a linear chain of nodes.
Undo the break of the blue edge, restoring the insertion point of x.
Let's begin with how to undo an unzip operation. This, fortunately, isn't too bad. We can identify the chains of nodes that we made with the unzip operation when we inserted x into the zip tree fairly easily - we simply look at the left and right children of x, then move, respectively, purely to the left and purely to the right.
Now, we know that these nodes used to be linked together in a chain. What order do we reassemble them into? As an example, take a look a this part of a zip tree, where we want to remove 53. The chains to the left and right of 53 are highlighted:
If we look at the nodes making up the left and right chains, we can see that there's only one way to reassemble them. The topmost node of the reassembled chain must be 67, since it has rank 3 and will outrank all other items. After that, the next node must be 41, because it's the smaller of the rank-2 elements and elements with the same rank have smaller items on top. By repeating this process, we can reconstruct the chain of nodes, as shown here, simply by using the rules for how zip trees have to be structured:
This operation, which interleaves two chains together into one, is called zipping.
To summarize, here's how a deletion works:
Deleting a node x from a zip tree works as follows:
Find the node x in the tree.
Perform a zip of its left and right subtrees. Specifically:
Maintain "lhs" and "rhs" pointers, initially to the left and right subtrees.
While both those pointers aren't null:
If lhs has a higher rank than rhs, make lhs's right child rhs, then advance lhs to what used to be lhs's right child.
Otherwise, make rhs's left child lhs, then advance rhs to point to what used to be rhs's left child.
Rewire x's parent to point to the result of the zip operation rather than x.
More to Explore
To recap our main points: we saw how to represent a skiplist as a BST by using the idea of ranks. That gave rise to the zip tree, which uses ranking rules to determine parent/child relationships. Those rules are maintained using the zip and unzip operations, hence the name.
Doing a full analysis of a zip list is basically done by reasoning by analogy to a skiplist. We can show, for example, that the expected runtime of an insertion or deletion is O(log n) by pointing at the equivalent skiplist and noting that the time complexity of the equivalent operations there are O(log n). And we can similary show that these aren't just expected time bounds, but expected time bounds with a high probability of occurring.
There's a question of how to actually store the information needed to maintain a zip tree. One option would be to simply write the rank of each item down in the nodes themselves. That works, though since ranks are very unlikely to exceed O(log n) due to the nature of geometric random variables, that would waste a lot of space. Another alternative would be to use a hash function on node addresses to generate a random, uniformly-distributed integer in some range, then find the position of the most least-significant 1 bit to simulate our coin tosses. That increases the costs of insertions and deletions due to the overhead of computing the hash codes, but also decreases the space usage.
Zip trees aren't the first data structure to map skiplists and BSTs together. Dean and Jones developed an alternative presentation of this idea in 2007. There's also another way to exploit this connection. Here, we started with a randomized skiplist, and used it to derive a randomized BST. But we can run this in reverse as well - we can start with a deterministic balanced BST and use that to derive a deterministic skiplist. Munro, Papadakis, and Sedgewick found a way to do this by connecting 2-3-4 trees and skiplists.
And zip trees aren't the only randomized balanced BST. The treap was the first structure to do this, and with a little math you can show that treaps tend to have slightly lower expected heights than zip trees. The tradeoff, though, is that you need more random bits per node than in a zip tree.
Hope this helps!
I'm implementing the Bentley-Ottmann algorithm
to find the set of segment intersection points,
unfortunately I didn't understand some things.
For example :
how can I get the neighbours of the segment Sj in the image.
I'm using a balanced binary search tree for the sweepLine status, but we store the segments in the leaves, after reading this wikipedia article I didn't find an explanation for this operation.
From the reference book (de Berg & al.: "Computational Geometry", ill. at p.25):
Suppose we search in T for the segment immediately to the left of some point p that lies on the sweep line.
At each internal node v we test whether p lies left or right of the segment stored at v.
Depending on the outcome we descend to the left or right subtree of v,
eventually ending up in a leaf.
Either this leaf, or the leaf immediately to the left of it, stores the segment we are searching for.
For my example if I follow this I will arrive at the leaf Sj but I will know just the leaf to the left i.e. Sk, how can I get Si?
Edit
I found this discussion that looks like my problem, unfortunately there are no answers about how can I implement some operations in such data structure.
The operation are:
inserting a node in such data structure.
deleting a node.
swapping two nodes.
searching for neighbours' node.
I know how to implement these operations in a balanced binary search tree when we store data too in internal node, but with this type of AVL I don't know if it is the same thing.
Thank you
I stumbled upon the same problem when reading Computational Geometry from DeBerg (see p. 25 for the quote and the image). My understanding is the following:
say you need the right neighbor of a segment S which is in the tree. If you store data in the nodes, the pseudo code is:
locate node S
if S has a right subtree:
return the left-most node of the right subtree of S
else if S is in the left sub-tree of any ancestor:
return the lowest/nearest such ancestor
else
return not found
If you store the data in the leaves, the pseudo-code becomes:
let p the point of S currently on the sweep line
let n the segment at the root of the tree
while n != null && n is not a leaf:
if n = S:
n = right child of S
else:
determine if p is on the right or left of n
update n accordingly (normal descent)
In the end, either n is null and it means there is no right neighbor, or n points to the proper leaf.
The same logic applies for the left neighbor.
Same as you, I have met the same problem while reading the de Berg & al.: "Computational Geometry". But I think The C++ Standard Template Library (STL) have an implantation called "map" which can do the job.
You just need to define some personalized class for line segment and event points and their comparison functions. Then, use std::map to build the tree and access the neighboring element using map.find() to get and iterator, and use iterator to gain access to the two neighbor element.
I am learning Left-Lean-Red-Black tree, from Prof.Robert Sedgewick
http://www.cs.princeton.edu/~rs/talks/LLRB/LLRB.pdf
http://www.cs.princeton.edu/~rs/talks/LLRB/RedBlack.pdf
While I got to understand the insert of the 2-3 tree and the LLRB, I have spent totally like 40 hours now for 2 weeks and I still can't get the deletion of the LLRB.
Can anyone really explain the deletion of LLRB to me?
Ok I am going to try this, and maybe the other good people of SO can help out. You know how one way of thinking of red nodes is as indicators of
where there there imbalance/new nodes in the tree, and
how much imbalance there is.
This is why all new nodes are red. When the nodes (locally) balance out, they undergo a color flip, and the redness is passed up to the parent, and now the parent may look imbalanced relative to its sibling.
As an illustration, consider a situation where you are adding nodes from larger to smaller. You start with node Z which is now root and is black. You add node Y, which is red and is a left child of Z. You add a red X as a child of Z, but now you have two successive reds, so you rotate right, recolor, and you have a balanced, all black (no imbalance/"new nodes"!) tree rooted at Y [first drawing]. Now you add W and V, in that order. At first they are both red [second drawing], but immediately V/X/W are rotated right, and color flipped, so that only X is red [third drawing]. This is important: X being red indicates that left subtree of Y is unbalanced by 2 nodes, or, in other words, there are two new nodes in the left subtree. So the height of the red links is the count of new, potentially unbalanced nodes: there are 2^height of new nodes in the red subtree.
Note how when adding nodes, the redness is always passed up: in color flip, two red children become black (=locally balanced) while coloring their parent red. Essentially what the deletion does, is reverse this process. Just like a new node is red, we always also want to delete a red node. If the node isn't red, then we want to make it red first. This can be done by a color flip (incidentally, this is why color flip in the code on page 3 is actually color-neutral). So if the child we want to delete is black, we can make it red by color-flipping its parent. Now the child is guaranteed to be red.
The next problem to deal with is the fact that when we start the deletion we don't know if the target node to be deleted is red or not. One strategy would be to find out first. However, according to my reading of your first reference, the strategy chosen there is to ensure that the deleted node can be made red, by "pushing" a red node down in front of the search node as we are searching down the tree for the node to be deleted. This may create unnecessary red nodes that fixUp() procedure will resolve on the way back up the tree. fixUp() presumably maintains the usual LLRBT invariants: "no successive red nodes" and "no right red nodes."
Not sure if that helps, or if we need to get into more detailed examination of code.
There is an interesting comment about the Sedgwich implementation and in particular its delete method from a Harvard Comp Sci professor. Left-Leaning Red-Black Trees Considered Harmful was written in 2013 (the Sedgwich pdf you referenced above is dated 2008):
Tricky writing
Sedgewick’s paper is tricky. As of 2013, the insert section presents 2–3–4 trees as the default and describes 2–3 trees as a variant. The delete implementation, however, only works for 2–3 trees. If you implement the default variant of insert and the only variant of delete, your tree won’t work. The text doesn’t highlight the switch from 2–3–4 to 2–3: not kind.
The most recent version I could find of the Sedgwich code, which contains a 2-3 implementation, is dated April 2014. It is on his Algorithms book site at RedBlackBST.java
Follow the next strategy to delete an arbitrary node in a LLRB tree which is not in a leaf:
Transform a LLRB tree to a 2-3-4 tree (we do not need to transform the whole tree, only a part of the tree).
Replace the value of the node (which we want to delete) its successor.
Delete its successor.
Fix the tree (recover balance, see the book "Algorithms 4th edition" on the pages 435, 436).
If a value in a leaf then we do not need to use a successor to replase this value, but we still need to transform the current tree to 2-3-4 tree to delete this value.
The slide on the page 20 of this presentation https://algs4.cs.princeton.edu/lectures/keynote/33BalancedSearchTrees.pdf and the book "Algorithms 4th edition" on the page 437 are a key. They show how a LLRB tree transformations into a 2-3 tree. In the book "Algorithms 4th edition" on the page 442 https://books.google.com/books?id=MTpsAQAAQBAJ&pg=PA442 is an algorithm of transformation for trees.
For example, open the page 54 of the presentation https://www.cs.princeton.edu/~rs/talks/LLRB/08Dagstuhl/RedBlack.pdf. The node H has children D and L. According to the algorithm on the page 442 we transform these three nodes into the 4-node of a 2-3-4 tree. Then the node D has children B and F we also transform these nodes into a node of 2-3-4 tree. Then the node B has children A and C we also transform these nodes into a node of 2-3-4 tree. And finally we need to delete A. After deletion we need to recover balance. We move up through the tree and we restore balance of the tree (according to rules, see the book "Algorithms 4th edition" on the pages 435, 436). If you need to delete the node D (the same tree on the page 54). You need the same transformations and need to replace the value of the node D on the value of the node E and delete the node E (because it is a successor of D).
I have this tree which, for each node, has exactly 10 childnodes (0-9). Each node has some associated data (say, for example, a name and a tag and a color) which, I guess, isn't important for this question. Each of the childnodes has exactly 10 childnodes. A node can be null (which 'ends' the branch') or contain another node.
To visualize what I'm talking about I made this diagram (fear my paintz0r skillz!):
A black box is a null-node. A white box is a node which contains data and childnodes. As you can see, even the root, each node has exactly 10 childnodes. Because of simplicity and to keep the diagram sane I have drawn some nodes very tiny but you can imagine these tiny nodes being the same.
This structure allows me to traverse a path consisting of digits very quickly: a path of 47352 would lead me down the "orange path" to the final destination; 4->7->3->5 where the final 2 cannot be resolved because that last one is a null-node (although colored red) and contains no childnodes.
My question is pretty simple actually: what is this kind of tree called? I have gone through all trees on Wikipedia's Tree (data structure) lemma and the closest I (think I) could get is the Octree and/or K-ary tree. Along those lines of reasoning my tree would be called a Dectree, Decitree, 10-ary tree or 10-way tree or something. But there might be a better name for this. So: anyone?
K-ary tree with K=10
In graph theory, a k-ary tree is a rooted tree in which each node has
no more than k children
It is also sometimes known as a k-way tree, an N-ary tree, or an M-ary
tree. A binary tree is the special case where k=2.
This is something like B-Tree.
I have to implement Fortunes algorithm for constructing Voronoi diagrams.
Important part of the algorithm is a data structure called "Beach Line Data Structure".
It is a binary balanced tree, similar to AVL, but different in a way that data is stored only on the leafs (there are other differences, but are unimportant for the question).
I am not sure how to implement it. Obviously using AVL "as is" will not work because when balancing AVL tree leaf node can become inner node and vice versa.
I also tried to look at some other known data structures at wikipedia, but none suits the needs.
I have seen some implementations that do this with a linked list, but this is not good because searching linked list is O(n), and it needs to be O(log n) for the algorithm to be efficient.
The leaves indeed store (single) points and the inner nodes of the event structure (the "beach line tree") stores ordered tuples of points whose parabolas/arcs lie next to each other. If the parabola that point Pa forms lies to the left of the parabola formed by Pb (and these two parabola's intersect), the inner node stores the ordered tuple (Pa, Pb).
Obviously using AVL "as is" will not work because when balancing AVL tree leaf node can become inner node and vice versa.
If you're worried about storing different types of objects in the AVL tree, a simple scheme would be to store the leaves as tuples too. So don't store point Pj as a leaf, but store the tuple (Pj, Pj) instead. If Pj as a leaf disappears from the event tree (beach line), and its parent is (Pi, Pj), simply change the parent into (Pj, Pj), and of course its parent will also needs to be changed from (Pj, P?) to (Pi, P?) etc. Just as with a regular AVL tree: you walk up the tree and modify the inner nodes that need to be changed and/or re-balanced.
Note that a good implementation of the algorithm can't be easily written down in a SO answer (at least, not by me!). For a proper explanation of the entire algorithm, including a description of the data structures used by it, see Computational geometry: algorithms and applications by Mark de Berg et al.. Chapter 7 is devoted solely to Voronoi diagrams.