Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Van Emde Boas Trees

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Dynamic algorithms – F04.

Q4, March 29, 2004

van Emde Boas trees.


Consider the datatype ordered dictionary. In addition to the usual dictio-
nary operations (insert, delete, lookup) it supports operations for finding the
immediately larger entry and possibly also the immediately smaller entry
compared to a given element (findnext, findprevious). Of course, the extra
operations require keys to be taken from an ordered domain.
In a standard binary search tree implementation of an ordered dictionary
some operations take time O(log n). van Emde Boas observed that if keys
are restricted to the set {1, 2, 3, ..., n} then all operations can be done in time
O(log log n) and space O(n) using a special data structure that has become
known under the name of van Emde Boas trees [3, 2].
This note describes Emde Boas trees with an application, namely the im-
plementation of Union-Split-Find on intervals. Other applications are given
in the problem set.
We start by considering a restricted datatype defining only three opera-
tions:

• Name: Tree(n)

• State: T is subset of {1, 2, 3, . . . , n}.

• Operations:
insert(x) Add x to T .
delete(x) Remove x from T .
findsucc(x) Return the smallest element in T that is ≥ x.

Consider possible implementations. If the set is represented as a bit array,


insert and delete take time O(1), but findsucc is very inefficient; in the
worst case, it may take time O(n). Using red-black trees allows a time bound
of O(log n) for all operations.
Something better is possible, if we exploit that the underlying key universe
is the set {1, 2, 3, . . . , n}. We can
√ make a recursive representation,
√ where
Tree(n) distribute elements over n distinct versions
√ of Tree( n). The
√latter
trees are named bottom[i] for i = 0, 1, 2, . . . , n − 1. An element i = a n + b
is represented by letting bottom[a] have the entry b.

1
Note that the translation from i to a, b (and conversely) is simple and
expedient, since a and b is the most significant half of bits in the
√ binary rep-
resentation of i, respectively the least significant half of bits ( n is rounded
to a power of 2). √
Tree(n) uses an additional version of Tree( n) called top. top contains
a if and only if bottom[a] is nonempty. It is now possible to implement
an operation of Tree(n)
√ in constant time plus the time for calling a single
operation in Tree( n) (implying O(log log n) time per operation).
This method may be viewed as binary search on the path from the root
to a leaf in a balanced binary tree. Initially, we spend constant time deciding
whether to call a top operation (upper half of path) or a bottom operation
(lower half of path). This is followed by the recursive call.
In this way, findsucc can be done in time O(log log n). It requires more
sophistication to obtain the same bound for delete and insert, but a partial
implementation is sketched in datastructure 1.

Data structure 1 Tree(n) in version 1


State: T is a subset of {1, 2, 3, ..., n} represented by
size: integer (number of elements in T )
max,min: √ integer (largest and smallest element of T )
top: Tree( n) √ √
bottom: array [1.. n] of Tree( √ n)
Operation findsucc(i) {i = a n √ + b}
if bottom[a].max ≥ b → j := a n+ bottom[a].findsucc(b)

[] else → c := top.findsucc(a + 1); j := c n+ bottom[c].min
fi
return j; √
Operation insert(i) {i = a n + b}
if bottom[a].size = 0 → top.insert(a) fi
bottom[a].insert(b);
update size,max,min √
Operation delete(i) {i = a n + b}
bottom[a].delete(b);
if bottom[a].size = 0 → top.delete(a) fi
update size,max,min

Note that max,min may be updated in time O(1) when using updated
values of max,min from top and bottom.
In the specific case when we insert the first element in a bottom tree or
delete the last element from a bottom tree, we need in addition to insert,
respectively remove, an element from the top tree. This makes a worst case

2

recurrence for time usage of T (n) = 2T ( n) + O(1) with the solution T (n) =
O(log n). To obtain a better result, we must modify the implementation to
allow insertion of the first element into an empty tree in constant time and
to allow deletion of the last element from a tree in constant time, i.e. we
should avoid recursive calls. It is possible to modify the implementation in
the suggested manner and still make it work. In every nonempty tree we keep
one element (say the minimum) out of the recursive structure. The√ details are
shown in datastructure 2, and it leads to the recurrence T (n) = T ( n)+O(1)
that has the solution T (n) = O(log log n).

Data structure 2 Tree(n) in version 2


State: Identical to version 1
Operation findsucc(i)
identical to findsucc from version 1 with the following statement added
just before the return statement:
if i ≤ min < j → j := min fi
Operation insert(i)
if size = 0 → update size,max,min
[] size > 0 →
if min > i → swap(min,i) fi
insert(i) from version 1
fi
Operation delete(i)
if size ≤ 1 → update size,max,min
[] size ≥ 2 → √
if min = i → i := min := top.min· n+bottom[top.min].min fi
delete(i) from version 1
fi

Note that we might also have kept max out of the recursive structure by
a straightforward generalisation of data structure 2.
The implementation is described recursively, but for small n the set should
be represented directly in a bitstring or a red-black tree.
In practice we start with a welldefined tree structure that is either empty
or completely full. Either possibility can be handled in time O(1) using lazy
initialization.
We have restricted the description to the operations insert, delete and
findsucc, but in great detail. These operations are fundamental, and it is
easy to implement additional operations such as deletemin being equal to
delete(findsucc(1)) and decreasekey that may be implemented by com-
bining delete and insert.

3
This ordered dictionary has several applications (see the problem set).
We describe one of them here, namely Union-Find-Split on intervals.

Union-Split-Find on intervals.
The usual union-find datastructure is very efficient for representing a parti-
tioning of a set into classes with operations for uniting two classes and for
getting the name of the class containing some given element. If one wants
an extra operation for splitting a class into two classes, it is necessary to
define more precisely what such a split operation should do. A class with
k elements may be split in 2k − 2 different ways. For a survey of different
union-split-find problems and known solutions, see [1].
We present here a very efficient solution (time O(log log n) per operation)
for a variant of union-split-find on intervals, ie. we consider the following
datatype:

Union-Split-Find:
• Name: Interval(n)

• State: T is a partitioning of {1, 2, 3, . . . , n} into intervals.

• Operations:
find(x) returns the name of the interval containing x.
union(x): Unites the interval containing x with the immediately fol-
lowing interval.
split(x): The interval I containing x is split into two intervals I ∩[1; x]
and I ∩ [x + 1; n].

If an interval is represented by its rightmost element, then the partitioning


[1; x1 ], [x1 + 1; x2 ], . . . , [xk + 1; n] can be represented uniquely by the set S =
{x1 , x2 , ..., xk , n}. Operation find(x) must return the smallest element in
S that is greater than (or equal to) x, while operations union and split
removes (respectively inserts) an element from (into) S. Implementation of
union-split-find on intervals using van Emde Boas trees is straightforward
using this strategy.
Note that the representation is reversible: We may implement an ordered
dictionary by means of a union-split-find data structure. The two datastruc-
tures are in fact equivalent.

4
References
[1] Zvi Galil and Giuseppe F. Italiano. Data structures and algorithms for disjoint
set union problems. ACM Computing Surveys (CSUR) 23(3) (1991), 319–344.

[2] P. van Emde Boas. Preserving order in a forest in less than logarithmic time
and linear space. Inform. Process. Lett. 6(3) (1977), 80–82.

[3] P. van Emde Boas, R. Kaas, and E. Zijlstra. Design and implementation of
an efficient priority queue. Math. Systems Theory 10(2) (1976/77), 99–127.
Sixteenth Annual Symposium on Foundations of Computer Science (Berkeley,
Calif., 1975), selected papers.

You might also like