TWAYKL 12-10
(1989, 11, 20)
Voronoi Trees and Applications
Hartmut Noltemeier
Department of Computer Science
University of Wuerzburg, Wuerzburg, W-Germany
Abstract
We present a new data structure called Voronoi
Tree which
supports the solution of a lot of proximity problems in general
quasi metric spaces vith efficiently computable distance (dis-
tortion) functions.
We analyse some structural properties, report experimental re~
sults showing that Voronoi Trees are a proper and very efficient
tool for the representation of proximity properties and the gene-
ration of suitable clusterings too.
We point to some appropriate applications as for example the re-
presentation of large sets of geometric objects, and show that.
Voronoi Trees can be generated in optimal time and height. We
conclude with some further directions of research and anplications.
L Introduction
The appropriate and efficient re-
presentation of proximity properties
in large sets of objects is a crucial
problem in @ wide range of applica-
tions.
Not only wellknown segmentation and
clustering problems as for example
in the field of
= partitioning a set of processes in
distributed systems
= physically clustering of data
= clustering a large set of rules in
tule based deductive systems
~ Finding adequate type hierarchies
in object oriented data base en-
vironments
eta te
or in the area of
= segmenting images and
~ clustering large scenes of geomet ric
objects (like points, edges, poly-
gons ete.) usually with respect to
pattern recognition, scene analy-
sing and image understanding purpo-
have to be supported by suitable
structures and methods.
But also numerous Layout and design
problems as
~ VLSI placement
- board design
= design of flexible manufacturing
systems (FMS) etc.
seriously can take advantage fromthese basic tools (see f.e. [2], [4])-
In general proximity properties may be
modeled (or frequently are given im-
plicitly and can be extracted) by
sone kind of distance or distortion
function.
In this peper ve will concentrate on
the problen how to suppart proximity
and clustering problems in quasi me~
tric spaces, although most of the
ideas can be adapted to much more
general cases.
A set E is called quasi metric if
there is a distance function
iE? + fe with the following pro-
perties (for any e, e' e'' € E):
(a) dve,e)=0
(b) dlee' )=d¢er se)
(c) dlese"") Sdlese" ede",
(reflexivity)
(synmetry)
)
(triangle inequality).
However it may happen that there are
some different elements e,e' € & vith
dte,e!
In section 2 ve shortly review bisec-
tor trees and introduce a new data
structure, called Voronoi Tree.
In section 3 ve are especially con-
cerned with Voronoi Trees for large
sets of geometric objects.
In section 4 we conclude by reporting
some experimental results and outline
some further topics of research and
directions of applications.
~70
2 Voronoi Trees
At first let us shortly review bi-
sector trees, vhich were introduced
by Kalantari, McDonaid [8] for normed
spaces, but can be used in quasi me-
tric spaces as well.
A finite set S={e,,.-.,e,} of ele-
ments of a (quasi)metric space £ can
be represented by a binary tree T
(called ‘bisector tree’ or ‘bs-tree*
for short) by the folloving insertion
pracedur
(1) each node of the actual tree 1,
(representing the actual set
3
at least one (p,) and at most two
epee ey)s OSign) contains
elements (p, and pp) of S,.
(2)
to ins
rt_anev element e,,) into
1, (yielding T,,)) start at the
root node of T, as follovs:
- if the actual node v contains
only one element then insert
€:41 into v
= if (otherwise) v contains two
elements py and py then recur-
sively insert e into the
itl
left subtree (which has root p,)
if dle; 41h, )