IOI Training Week 7

Advanced Data Structures

Tim Dumol

Contents
1 Range Minimum Query 1
1.1 Square-root (sqrt) Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Segment Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Self-Balancing Binary Search Trees 3

3 Bonus: More interesting data structures 6

4 Problems 6
4.1 Bonus Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

5 References 6

1 Range Minimum Query

This section rotates around a common problem:
Definition 1 (Range Minimum Query (RMQ)). Given an integer array of fixed length, answer a set of
queries of the form: “What is the minimum element in the range of the array from i to j?”. The contents
of the array may change between queries.
The naı̈ve solution for RMQ has no setup time, and O(n) query time. We can improve on this by
adding some setup time, and using some additional memory. We will discuss two approaches: square-root
decomposition, and segment trees.

1.1 Square-root (sqrt) Decomposition

√ √
The idea behind√ sqrt decomposition is simple: preprocess the array into n chunks
√ of size n each (thus
consuming O( n) extra memory), so that we can perform the query in O( n) time, by using the pre-
processed chunks to compute the minimum for the parts √of the range that have a full intersection with
the chunks, and then traversing the remaining at most 2 ( n − 1) elements uncovered by the chunks1 . To
elaborate, in code2 :
struct SqrtDecomp {
v e c t o r <int >∗ a r r ;
v e c t o r <int> chunks ;
int c h u n k s i z e ;
i n t n chunks ;
SqrtDecomp ( v e c t o r <int> const ∗ a r r ) : a r r ( a r r ) {
c h u n k s i z e = ( i n t ) s q r t ( a r r −> s i z e ( ) ) ;
n chunks = ( i n t ) c e i l ( a r r −> s i z e ( ) / ( double ) n chunks ) ;
chunks . r e s i z e ( n chunks ) ;
1 One can see that this can be extended to any associative operation.
2 code is untested, if it’s wrong, feel free to correct

1
1 RANGE MINIMUM QUERY 2

f o r ( i n t i = 0 ; i < n chunks ; ++i ) {

// cap computed i n advance t o a v o i d r e c o m p u t i n g
const i n t cap = min ( n chunks ∗ ( i +1) , a r r −> s i z e ( ) ) ;
// a s s u m p t i o n : a l l i n p u t v a l u e s l e s s than 1 << 30
chunks [ i ] = 1 << 3 0 ;
f o r ( i n t j = i ∗ c h u n k s i z e ; j < cap ; ++j ) {
chunks [ i ] = min ( a r r [ j ] , chunks [ i ] ) ;
}
}
}

// end i s e x c l u s i v e
i n t query ( i n t begin , i n t end ) {
i n t l e f t = ( i n t ) c e i l ( b e g i n / ( double ) c h u n k s i z e ) ;
i n t r i g h t = ( i n t ) f l o o r ( end / ( double ) c h u n k s i z e ) ;
i n t ans = 1 << 3 0 ;
i f ( l e f t <= r i g h t ) {
ans = min ( ans , m i n e l e m e n t ( chunks . b e g i n ( ) + l e f t , chunks . b e g i n ( ) + r i g h t ) ) ;
}
i f ( s t a r t % c h u n k s i z e != 0 ) {
ans = min ( ans , m i n e l e m e n t ( a r r −>b e g i n ( ) + begin , a r r −>b e g i n ( ) + ⤦
Ç chunk size ∗ l e f t ) ) ;
}
i f ( end % c h u n k s i z e != 0 ) {
ans = min ( ans , m i n e l e m e n t ( a r r −>b e g i n ( ) + c h u n k s i z e ∗ r i g h t , a r r −>b e g i n ( ) + ⤦
Ç end ) ) ;
}
return ans ;
}
};

The code to update the sqrt decomposition is an exercise left to the reader (you don’t need to submit it).

1.2 Segment Trees

√
But a query time O( n) is still pretty slow. Can we do faster? The answer is yes. We can get O(log(n))
query time with O(n) extra memory, by using a segment tree.
Definition 2 (Segment Tree). A segment tree over an array of length n (for simplicity, let’s say it’s a power
of two – extending to a non-power of two is an exercise for the reader) is a balanced binary3 tree with n
leaves, each corresponding to an element in the array. Each internal vertex of the segment tree has a value
corresponding to the minimum of all vertices under its subtree.
Since a segment tree is a complete binary tree, it can be represented similarly as a heap, using only a
single-dimensional integer array. Furthermore, a segment tree has 2n − 1 vertices, and can be constructed in
O(n) time. For querying, we recursively traverse the segment tree, stopping when the segment covered by a
vertex is wholly included in the query range. It can be shown that this results in O(log(n)) time4 .
struct ST {
typedef i n t e l t y p e ;
v e c t o r <e l t y p e > t r e e ;
int n ;
ST( v e c t o r <e l t y p e > const& a r r ) {
n = arr . si ze () ;
t r e e . r e s i z e (2∗ arr . s i z e ( ) ) ;
3 technically you can use any arity, but for simplicity let’s say binary
4 https://cs.stackexchange.com/questions/37669/time-complexity-proof-for-segment-tree-implementation-of-the-ranged-sum-proble
2 SELF-BALANCING BINARY SEARCH TREES 3

build (0 , arr , 0 , arr . s i z e ( ) ) ;

}

i n l i n e i n t l e f t ( i n t i d x ) const { return 2∗ i d x + 1 ; }
i n l i n e i n t r i g h t ( i n t i d x ) const { return 2∗ i d x + 2 ; }

// end i s e x c l u s i v e l , as u s u a l
i n t b u i l d ( i n t idx , v e c t o r <e l t y p e > const& a r r , i n t s t a r t , i n t end ) {
i f ( s t a r t + 1 == end ) {
return ( t r e e [ i d x ] = a r r [ s t a r t ] ) ;
}
const i n t mid = ( s t a r t + end ) / 2 ;
return ( t r e e [ i d x ] = min ( b u i l d ( l e f t ( i d x ) , a r r , s t a r t , mid ) , b u i l d ( r i g h t ( i d x ) , ⤦
Ç a r r , mid , end ) ) ) ;
}

i n t query ( i n t s t a r t , i n t end ) {
return query ( s t a r t , end , 0 , 0 , n ) ;
}

i n t query ( i n t s t a r t , i n t end , i n t idx , i n t t r e e s t a r t , i n t t r e e e n d ) {

i f ( s t a r t <= t r e e s t a r t && t r e e e n d <= end ) {
return t r e e [ i d x ] ;
} else {
const i n t mid = ( t r e e s t a r t + t r e e e n d ) / 2 ;
return min ( query ( s t a r t , end , l e f t ( i d x ) , t r e e s t a r t , mid ) , query ( s t a r t , ⤦
Ç end , r i g h t ( i d x ) , mid , t r e e e n d ) ) ;
}
}
};

The code to update a single element is an exercise left to the reader (you don’t need to submit it).
Furthermore, through the usage of lazy propagation, you can make it so that you can update a range of
elements in the segment tree (also an exercise left to the reader – hint: you need to only add a single piece
of additional information, if your update is adding an integer to a given range).
Similar to sqrt decomposition, you can adapt the segment tree to any associative operation (e.g., range
sum query).

1.3 Notes
Coincidentally, RMQ turns out to be deeply related to another (somewhat rarer, but also standard) problem,
LCA:
Definition 3 (Lowest Common Ancestor (LCA)). Given a rooted tree, T, answer a set of queries of the
form: “What is the vertex of T that is farthest from the root, that is an ancestor of both vertices x and y?”.
The naı̈ve solution for LCA has no setup time, and O(n) query time, and it turns out you can preprocess
LCA into an RMQ problem, and vice versa. You can also use square-root decomposition for LCA, although
of a different form. To learn more about this (and asymptotically better solutions to RMQ), check out the
RMQ and LCA tutorial on Topcoder (c.f. references).

2 Self-Balancing Binary Search Trees

Recall that most library implementations of the map and set datastructures use a self-balancing binary search
tree (usually a Red-Black tree), allowing query, delete, and insert operations in amortized O(log(n)) time.
Usually, this is good enough for our purposes, but sometimes we need to augment each vertex with some
additional information in order to enable a special kind of query not supported by library implementations.
2 SELF-BALANCING BINARY SEARCH TREES 4

For example, you may have a dynamically changing ordered list of integers, and have to find the index of an
arbitrary integer in the list.
Now, the most commonly used trees in library code are AVL trees and Red-Black trees, because of their
good asymptotic characteristics. However, they are a pain to code, and have a lot of edge cases. Thus,
for competitive programming, we usually implement simpler trees: either scapegoat trees or treaps. In this
discussion, we’ll focus on treaps (because the author considers them easier to implement and understand)5 .
A treap is a binary search tree augmented with a randomized priority value on each of its vertices, and
kept balanced by maintaining a heap (priority queue) structure on its vertices (i.e., each vertex should have a
smaller priority than its children). Because a random heap is balanced, a treap is probabilistically balanced.
The hard part in implementing a treap is in maintaining the heap structure, as one will have to rotate
vertices that violate the heap structure, while maintaining the BST structure.
struct TreapNode {
i n t s t a r t , end ;
int p r i o r i t y ;
TreapNode ∗ k i d s [ 2 ] ;
TreapNode ∗ p a r e n t ;
TreapNode ( ) {}
void i n i t ( i n t s t a r t , i n t end , TreapNode ∗ p a r e n t ) ;
void u p d a t e a u g ( ) {
// w h a t e v e r a u g m e n t a t i o n you need
}
};

// we c a c h e t h e o b j e c t s so t h a t we don ’ t need t o d y n a m i c a l l y
// a l l o c a t e o b j e c t s i n heap ( t h i s i s a l s o known as arena
// a l l o c a t i o n )
TreapNode c a c h e [ 3 0 0 0 0 2 4 ] ;
int c c t r ;
void TreapNode : : i n i t ( i n t s t a r t , i n t end , TreapNode ∗ p a r e n t ) {
k i d s [ 0 ] = k i d s [ 1 ] = NULL;
this −>p a r e n t = p a r e n t ;
this −> s t a r t = s t a r t ;
this −>end = end ;
p r i o r i t y = rand ( ) ;
}

struct Treap {
TreapNode ∗ r o o t ;
Treap ( ) : r o o t (NULL) {}
// d i r : 0 i s l e f t , 1 i s r i g h t
void r o t a t e ( TreapNode ∗ node , i n t d i r ) {
TreapNode ∗ r k i d = node−>k i d s [ d i r ˆ 1 ] ;
i f ( node−>p a r e n t ) {
i f ( node == node−>parent −>k i d s [ 0 ] ) {
node−>parent −>k i d s [ 0 ] = r k i d ;
} else {
node−>parent −>k i d s [ 1 ] = r k i d ;
}
} else {
root = rkid ;
}
r k i d −>p a r e n t = node−>p a r e n t ;
node−>k i d s [ d i r ˆ 1 ] = r k i d −>k i d s [ d i r ] ;
i f ( node−>k i d s [ d i r ˆ 1 ] ) {
5 But
briefly: scapegoat trees remain balanced by maintaining some statistics on their subtrees, and if a subtree is sufficiently
“imbalanced”, it completely reconstructs that subtree. This rebalancing is done rarely enough that operations on a scapegoat
tree remain O(log(n)) time.
2 SELF-BALANCING BINARY SEARCH TREES 5

node−>k i d s [ d i r ˆ1]−> p a r e n t = node ;

}
// f i n a l l y , t r a n s p l a n t o r i g t o new p a r e n t
r k i d −>k i d s [ d i r ] = node ;
node−>p a r e n t = r k i d ;

// u p d a t e a u g m e n t a t i o n
node−>u p d a t e a u g ( ) ;
r k i d −>u p d a t e a u g ( ) ;
}
void add ( i n t s t a r t , i n t end ) {
i f ( r o o t == NULL) {
r o o t = &c a c h e [ c c t r ++];
r o o t −> i n i t ( s t a r t , end , NULL) ;
return ;
}
// l o o k f o r node t o i n s e r t a t
TreapNode ∗p = r o o t ;
TreapNode ∗ node = NULL;
while ( true ) {
i f ( s t a r t < p−> s t a r t ) {
i f ( ! p−>k i d s [ 0 ] ) {
node = p−>k i d s [ 0 ] = &c a c h e [ c c t r ++];
break ;
} e l s e p = p−>k i d s [ 0 ] ;
} else {
i f ( ! p−>k i d s [ 1 ] ) {
node = p−>k i d s [ 1 ] = &c a c h e [ c c t r ++];
break ;
} e l s e p = p−>k i d s [ 1 ] ;
}
}
node−> i n i t ( s t a r t , end , p ) ;
// m a i n t a i n a u g m e n t a t i o n
TreapNode ∗ p2 = p ;
while ( p2 != NULL) {
p2−>u p d a t e a u g ( ) ;
p2 = p2−>p a r e n t ;
}

// r o t a t e
while ( node−>p a r e n t && node−> p r i o r i t y > node−>parent −> p r i o r i t y ) {
i f ( node == node−>parent −>k i d s [ 0 ] ) {
r o t a t e ( node−>parent , 1 ) ;
} else {
r o t a t e ( node−>parent , 0 ) ;
}
}
}
i n t s e a r c h ( TreapNode ∗p , i n t s t a r t , i n t end ) {
// problem − d e p e n d e n t
}
i n t s e a r c h ( i n t s t a r t , i n t end ) {
return s e a r c h ( r o o t , s t a r t , end ) ;
}
};
3 BONUS: MORE INTERESTING DATA STRUCTURES 6

3 Bonus: More interesting data structures

This is just a listing of interesting data structures that you may want to look into:

• Binary Index Tree (BIT) (aka Fenwick Tree) https://www.topcoder.com/community/data-science/

data-science-tutorials/binary-indexed-trees/ – used to get prefix sums; functionality is a sub-
set of Segment Tree, but has faster to type implementation
• k-d tree https://en.wikipedia.org/wiki/K-d_tree – allows you to query the nearest neighbor of a
point among a dynamic set of points in k-dimensional space, in O(log(n)) time. (out of IOI scope)
• Skip list https://en.wikipedia.org/wiki/Skip_list – randomized data structure with performance
characteristics similar to a balanced binary search tree (very rare usage in competitive programming)

4 Problems
1. Reverse Prime (https://uva.onlinejudge.org/index.php?option=com_onlinejudge&Itemid=8&page=
show_problem&problem=2657)
2. Frequent Values (https://uva.onlinejudge.org/index.php?option=onlinejudge&page=show_problem&
problem=2176)
3. Census (https://uva.onlinejudge.org/index.php?option=com_onlinejudge&Itemid=8&page=show_
problem&problem=2272)
4. Can you answer these queries VIII (http://www.spoj.com/problems/GSS8/)
5. Permutations (https://uva.onlinejudge.org/index.php?option=onlinejudge&page=show_problem&
problem=2520)

6. Ahoy, Pirates! (https://uva.onlinejudge.org/index.php?option=com_onlinejudge&Itemid=8&

category=24&page=show_problem&problem=2397)
7. XOR on Segment (http://codeforces.com/problemset/problem/242/E)
8. Robotic Sort http://www.spoj.com/problems/CERC07S/

4.1 Bonus Problems

These problems are ungraded.

1. Alien Abduction (https://icpcarchive.ecs.baylor.edu/index.php?option=com_onlinejudge&Itemid=

8&category=559&page=show_problem&problem=4056)
2. Genetics (https://www.codechef.com/problems/GENETICS)

3. Little Elephant and Tree (http://codeforces.com/problemset/problem/258/E)

5 References
1. RMQ and LCA tutorial https://www.topcoder.com/community/data-science/data-science-tutorials/
range-minimum-query-and-lowest-common-ancestor/