Voronoi Diagrams - A Survey of A Fundamental Geometric Data Structure

Voronoi Diagrams — A Survey of a Fundamental Geometric
Data Structure
FRANZ AURENHAMMER
Institute fur Informationsverarbeitung Technische Universitat Graz, Sch iet!stattgasse 4a, Austria
This paper presents a survey of the Voronoi diagram, one of the most fundamental data
structures in computational geometry. It demonstrates the importance and usefulness
of the Voronoi diagram in a wide variety of fields inside and outside computer science
and surveys the history of its development. The paper puts particular emphasis on the
unified exposition of its mathematical and algorithmic properties. Finally, the paper
provides the first comprehensive bibliography on Voronoi diagrams and related
structures.
Categories and Subject Descriptors: F.2.2 [Analysis of Algorithms and Problem

Complexity]: Nonnumerical Algorithms and Problems–geometrical problems and
computations; G. 2.1 [Discrete Mathematics]: Combinatorics— combinatorial
algorithms; I. 3.5 [Computer Graphics]: Computational Geometry and Object
Modeling—geometric algorithms, languages, and systems
General Terms: Algorithms, Theory
Additional Key Words and Phrases: Cell complex, clustering, combinatorial complexity,
convex hull, crystal structure, divide-and-conquer, geometric data structure, growth
model, higher dimensional embedding, hyperplane arrangement, k-set, motion
planning, neighbor searching, object modeling, plane-sweep, proximity, randomized
insertion, spanning tree, triangulation
INTRODUCTION [19841 and to the textbooks by Preparata

and Shames [1985] and Edelsbrunner
Computational geometry is concerned [1987bl.)
with the design and analysis of algo- Readers familiar with the literature of
rithms for geometrical problems. In add- computational geometry will have no-
ition, other more practically oriented, ticed, especially in the last few years, an
areas of computer science— such as com- increasing interest in a geometrical con-
puter graphics, computer-aided design, struct called the Voronoi diagram. This
robotics, pattern recognition, and opera- trend can also be observed in combinato-
tions research—give rise to problems that rial geometry and in a considerable num-
inherently are geometrical. This is one ber of articles in natural science journals
reason computational geometry has at- that address the Voronoi diagram under
tracted enormous research interest in the different names specific to the respective
past decade and is a well-established area area. Given some number of points in the
today. (For standard sources, we refer to plane, their Voronoi diagram divides the
the survey article by Lee and Preparata plane according to the nearest-neighbor
Permission to copy without fee all or part of this material is granted provided that the copies are not made
or distributed for direct commercial advantage, the ACM copyright notice and the title of the publication
and its date appear, and notice is given that copying is by permission of the Association for Computing
Machinery. To copy otherwise, or to republish, requires a fee and/or specific permission.
@ 1991 ACM 0360-0300/91/0900-0345 $01.50
ACM Computing Surveys, Vol. 23, No. 3, September 1991

346 ● Franz A urenhammer
CONTENTS developed for the computer construction

and representation of Voronoi diagrams.
The intention of this survey is three-
fold: First, motivated by the fact that
INTRODUCTION Voronoi diagrams have been (reinvented
1. HISTORICAL PERSPECTIVE and studied fairly independently in the
11 The Natural Scientist’s Viewpoint
applied natural sciences, in mathemat -
1,2 The Mathematician’s Viewpoint
13 The Computer Scientist’s Viewpoint
its, and in computer science, it presents
2. ALGORITHMIC APPLICATIONS sketches of their historical development
2 1 Closest-Site Problems in these three areas. Second, it surveys
2,2 Placement and Motion Planning the literature on Voronoi diagrams and
23 Triangulating Sites
2,4 Connectivity Graphs for Sites
related structures, with particular em-
2,5 Clustering Point Sites phasis on the unified exposition of their
3, SELECTED TOPICS mathematical and computational proper-
3.1 The Geometry of Voronol Diagrams, Them ties and their applications in computer
Relatlon to Higher Dimensional ObJects.
science. Finally, it provides the first
3.2 The Topology of Planar Diagrams:
Divide-and-Conquer ConstructIon comprehensive bibliography on Voronoi
and its Variants. diagrams.
33 A Deformation of the Voronoi Diagram:
The Plane-Sweep Technique
Basic Properties of the Voronoi Diagram
REFERENCES
We begin with a description of elemen-

tary, though important, properties of the
rule: Each point is associated with Voronoi diagram that will suggest some
the region o~ the plane closest to it; feeling for this structure. We also intro-
(Figure 1). duce notation used throughout this pa-
Why do Voronoi diagrams receive so per. See also Preparata and Shames
much attention? What is special about [1985] or Edelsbrunner [1987] for sources
this easily defined and visualized con- on this material.
struct? It seems three main reasons are We first give a usual generic definition
responsible. First, Voronoi diagrams of the Voronoi diagram. Let S denote a
arise in nature in various situations. In- set of n points (called sites) in the plane.
deed, several natural processes can be For two distinct sites p, q e S, the domi-
used to define particular classes of nance of p over q is defined as the sub-
Voronoi diagrams. Human intuition is set of the plane being at least as close to
often guided by visual perception. If one p as to q. Formally,
sees an underlying structure, the whole
situation may be understood at a higher dom(p, q) = {x~l?zl~(x,p) <b(x, q)},
level. Second, Voronoi diagrams have in-
teresting and surprising mathematical
for 6 denoting the euclidean distance
properties; for instance, they are related
function. Clearly, dom( p, q) is a closed
to many well-known geometrical struc-
half plane bounded by the perpendicuhm
tures. Thifi has led several authors to
bisector of’ p and q. This bisector sepa-
believe that the Voronoi diagram is one
rates all points of the plane closer to p
of the most fundamental constructs de-
from those closer to q and will be termed
fined by a discrete set of points. Finally,
the separator of p and q. The region of a
Voronoi diagrams have proved to be a
site p e S is the portion of the plane ly-
powerful tool in solving seemingly un-
ing in all of the dominances of p over the
related computational problems and
remaining sites in S. Formally
therefore have increasingly attracted
the attention of computer scientists
in the last few years. Efficient and reg(p) = ,e~~{P1 dom(p, q).
reasonably simple techniques have been
ACM Computmg Surveys, Vol. 23, No. 3, September 1991

Voronoi Diagrams ● 347
should be observed that each vertex is

the center of a circle that passes through
at least three sites but encloses no site.
Although n sites give rise to
();=
0( n2) separators, only linearly many
separators contribute an edge to V(S).
This can be seen by viewing a Voronoi
diagram as a planar graph with n re-
gions and minimum vertex degree 3.
Each of the e edges has two vertices, and
each of the u vertices belongs to at least
three edges. Hence, 2 e >3 U. Euler’s re-
lation n + v – e ? 2 now implies e < 3n
– 6 and u s 2 n – 4, Thus, for example,
Figure 1. Voronoi diagram for eight sites in the the average number of edges of a region
plane. does not achieve six; there are less than
3 n edges, and each of them belongs to
exactly two of the n regions.
Since the regions are coming from in- The linear behavior of the size of the
tersecting n – 1 half planes, they are Voronoi diagram in the plane means that,
convex polygons. Thus the boundary of a roughly speaking, this structure is not
region consists of at most n – 1 edges much more complex than the underlying
(maximal open straight-line segments) configuration of sites. This is one of the
and vertices (their endpoints). Each point main reasons for the frequent use of
on an edge is equidistant from exactly Voronoi diagrams. A second reason is
two sites, and each vertex is equidistant that V(S) comprises the entire proximity
from at least three. As a consequence, information about S in an explicit and
the regions are edge to edge and vertex computationally useful manner. For ex-
to vertex, that is to say, they form a ample, its applicability to the important
polygonal partition of the plane. This post-office problem (see below) is based
partition is called the Voronoi diagram, on the trivial observation that a point x
V(S), of the finite point-set S (Figure 1). falls into the region of a site p if and
Note that a region, say reg( p), cannot only if p is closest to x among all sites in
be empty since it contains all points of S. Moreover, if site p is closest to site q,
the plane at least as close to p as to any then reg( p) and reg( q) share a common
other sites in S. In particular, p e reg( p). edge. This particularly implies that the
It follows that V(S) contains exactly n closest pair of sites in S gives rise to
regions. Some of them are necessarily some edge of V(S).
unbounded. They are defined by sites ly.
ing on the boundary of the convex hull of Applications in Computer Science
S because just for those sites there exist
To substantiate the usefulness of the
points arbitrarily far away but still clos-
Voronoi diagram in computer science, we
est.1 No vertices occur if and only if all briefly describe four situations where this
sites in S lie on a single straight line.
structure is used. The practicality y and
Such degenerate configurations also im-
diversity of these applications will
ply the existence of regions with only one
impart the appeal of Voronoi diagrams.
(unbounded) edge. Otherwise, three or
more edges meet at a common vertex. It
Associative File Searching
Consider some file of n two-attribute

lThe conuex hull of S is the smallest convex polY- records referring, for example, to latitude
gon that contains S. and longitude of a city or to age and
ACM Computmg Surveys, Vol. 23, No 3, September 1991

348 ● Franz Aurenhammer
income of a person. Suppose we are given

an additional target record R and we
quickly want to retrieve that record of
the file that matches R the best. If re-
trieval occurs frequently on the same file,
a supporting data structure is called for.
This associative file searching problem
was first posed by Knuth [1973] as a
two-dimensional generalization of the
usual (one-attribute) file-searching prob-
lem. It possibly is best known in its geo-
metric version under the name post-office
problem: Given a set S of n sites in the
plane (post offices), report a site closest Fig ure 2. Voronoi diagram for two dense clusters.
to a given query point q (the location of a
person). Note that there exists a trivial
0( n)-time solution by computing all n attribute data, similarity is reflected by
distances. There are various algorithms the proximity of sites in the plane. Prox-
in computational geometry that need imity, in turn, is revealed by properties
the post office problem as a subroutine of the Voronoi diagram for these sites.
[Preparata and Shames 1985]. For instance, dense subsets of sites give
Shames [1975a] first observed the rise to Voronoi regions of small area;
relevance of Voronoi diagrams to this regions of sites in a homogeneous cluster
problem. A site p is closest to q if and will have similar geometric shape; for
only if q falls into the region of p. In a clusters having orientation-sensitive
preprocessing step the Voronoi diagram density, the shapes of the regions will
of S is computed. To report a site closest exhibit a corresponding direction sensi-
to q, it now suffices to determine the tivity. Figure 2 gives an example. Ahuja
region that contains q. For this so-called
[19821 showed how to use these proper-
point-location problem, efficient solutions ties for clustering and matching sites.
have been developed by Kirkpatrick Voronoi diagrams support various
[19831, Edahiro et al. [1984], and Edels- clustering techniques used in practice.
brunner et al. [1986]. In particular, point What is required at any stage of
location in a Voronoi diagram with n the clustering process is often little
regions is supported in O(log n) time and more than the retrieval of the nearest-
0(n) storage overhead. This shows that neighbor sites of specified sites. They
the post-office problem can be solved by can be reported easily by examining the
means of Voronoi diagrams in logarith- edges of the regions of the specified sites.
mic query time and without increasing This applies to several hierarchical
the order of space, which is well known methods [Murtagh 1983], partitional
to be optimal already for usual file methods [Asano et al. 1988], and meth-
searching. ods involving cluster selection [Aggarwal
et al. 1989].
Cluster Analysis
Scheduling Record Accesses
The problem of automatically clustering
data arises frequently [Hartigan 19751. Consider a mass storage system repre-
Finding clusters means determining a sented by a two-dimensional array of grid
partition of the given set of data into points, each capable of storing one record.
subsets whose in-class members are simi- The time for the read/write head to move
lar and whose cross-class members are from point (x, y) to the point (u, U) is
dissimilar according to a predefined simi- proportional to I x – u I + I y – u 1, the
larity measure. In the case of two- distance between these points measured
ACM Computing Surveys, Vol 23, No 3, September 1991

.-
. .. . ..jgcJytj?JrJ”
Figure 3. Voronoi diagram and minimum span-
ning trees in the LI-metric,
in the L1-metric (i. e., Manhattan dis-

tance). One problem that arises in ac- Figure 4. Power diagram for seven circles
cessing batched requests is the following:

How is the head movement to be sched-
uled in order to retrieve the requested
records in minimum time (distance)? stop the system before a collision will
Finding an exact solution was shown have occurred.
to be NP-complete by Lee and Wong The robot system and the environment
[19801. A satisfactory approximate solu- of obstacles are usually modeled by
tion can be obtained quickly by means of polygonal objects. For the sake of prox-
Voronoi diagrams. In an initial step, the imity detection, collision-critical points
Voronoi diagram induced under the Ll- on the boundary of these objects may be
metric is computed in 0( n log n) time by circumscribed by circles whose radii cor-
taking the n requested grid points as respond to the tolerance threshold of the
sites. Now the minimum-Ll-length tree system. This reduces the problem to de-
spanned by the sites can be constructed tecting the intersection of circles. Colors
quickly: An edge of the tree can only may be assigned to the circles in order to
connect sites whose regions are neigh- distinguish between “harmless” inter-
bored in the Voronoi diagram [Hwang sections (among circles stemming from
1979; Lee and Wong 1980]. Figure 3 il- the same moving robot part or the same
lustrates this diagram (solid) and the obstacle) and others. Sharir [1985]
corresponding spanning tree (dashed). To pointed out that there will be no inter-
obtain a head movement whose length is section between circles of different colors
within a factor of 2 of the optimum, each if all the connected components formed
tree edge is traversed twice to access all by the circles are uncolored. Finding
the sites. connected components and checking col-
ors seem to involve an inspection of each
Collision Detection
pair of circles. The problem becomes easy,
however, if the power diagram (the
An important topic in controlling the mo- Voronoi diagram where the power lines
tion of robot systems is that of collision of the circles are taken as separators) is
detection. For a robot moving in an ob- available; see Aurenhammer [1988a]
stacle environment, one needs to deter- where this problem is solved in
mine collisions between moving robot O(n log n) time for n colored circles. The
subparts and stationary obstacles or be- crucial property that can be exploited
tween two separately moving subparts of is that the points of intersection of two
the robot. In particular, proximity detec- circles lie on their common power line
tion is important in order to be able to (Figure 4).

350 “ Franz Aurenhammer
1. HISTORICAL PERSPECTIVE ●
The history of’ Voronoi diagrams can be

traced back to the middle of the nine-
teenth century. Although the spectrum
of scientific disciplines that include in-
terest in Voronoi diagrams is broad, three
aspects have been emphasized:
(1) Their use in modeling natural phe-

nomena
(2) The investigation of their mathemat -
ical, in particular, geometrical, com-
binatorial, and stochastic properties
(3) Their computer construction and
representation
Accordingly, Voronoi diagrams are Figure 5. Cubic crystal structure.

useful in three respects: As a structure
per se that makes explicit natural pro-
cesses, as an auxiliary structure for in-
vestigating and calculating related The objective in this respect was the
mathematical objects, and as a data study of regions arising from regularly
structure for algorithmic problems that placed sites. Figure 5 illustrates some of
are inherently geometric. In all three the regions defined by 15 cubically or-
applications, efficient and practical algo- dered sites. In his comprehensive article
rithms for computing Voronoi diagrams on crystal structures, Niggli [1927] calls
are required. Since the first application them Wirkungsbereiche (domains of ac-
bears the initial seed for their investigation), a term widely used nowadays and
tion, let us consider the role that Voronoi used, with many synonyms, even in the
diagrams play in the natural sciences 1930s as demonstrated in the clarifying
first. note by Nowacki [1933]. Much effort has
been devoted to the fundamental crystal-
lographical question: Which types of do-
1.1 Natural Scientist’s Viewpoint
mains of action are capable of filling the
To visualize the appearance of a Voronoi plane or three space completely if only
diagram in nature, one could think of the congruent copies (and certain motions)
three-dimensional space being subdi- are to be used? Significant work was
vided into a manifold of crystals: From done, among others, by Niggli [1927],
several sites fixed in space, crystals start Delaunay (Delone) [1932], Nowacki
growing at the same rate in all directions [19761, and by Koch [19751. In fact, the
and without pushing apart but stopping question is a deep mathematical one, and
growth as they come into contact. The we will come back to it in the discussion
crystal emerging from each site in this of mathematical aspects of Voronoi
process is the region of space closer to diagrams.
that site than to all others. In other
words, the regions form a Voronoi
1. 1.2 Wigner-Seitz Zones
diagram in three-space.
A physiochemical system consists of a
1. 1.1 Don-rams of Action number of distinct sites, typically
molecules, ions, or atoms. Equilibrium
Most of the early work on Voronoi dia- and other properties of the system de-
grams was motivated by crystallography. pend on the spatial distribution of the

sites, which can be conveniently repre- Voronoi diagram. This follows from in-
sented by dividing the space between vestigations by Maxwell [1864]. Figure 7
them according to the nearest-neighbor shows sites for a spider web such that the
rule. The resulting Voronoi regions are distance between neighborhood sites cor-
often called Wigner-Seitz zones after responds to the tension of the edge they
Wigner and Seitz [1933], who were the define.
first to use them in metallurgy. Frank
and Kasper [1958] use these zones in the
1. 1.4 Thiessen Polygons
investigation of complex alloy structures;
Loeb [1970] bases his survey of cubic Geographical interest in Voronoi dia-
crystal structures on an intimately re - grams originates with the climatologist
lated concept, Allotting zones to sites is Thiessen [1911], who assigned proximity
of interest to molecular physicists, bio- polygons to observation sites in order to
chemists, material scientists, and physi- improve the estimation of precipitation
cal chemists. Via this approach Brostow averages over large areas. His method
and Sycotte [1975] estimate the coordina- was worked out in detail by Horton [19171
tion number of liquid argon, David and who proposed the name Thiessen poly-
David [1982] study certain solvation gons, crediting Thiessen with the idea.
structures, Brumberger and Goodisman As reported in Boots [1979], Thiessen
[19831 interpret the small-angle scatter- polygons play four roles in geographical
ing of catalysts, and Augenbaum and research: As models for spatial processes,
Peskin [1985] treat large-scale hydro- as nonparametric techniques in point-
dynamic codes—just to name a few. pattern analysis, as organizing struc-
tures for displaying spatial data, and for
calculating individual probabilities in
1.1.3 Johnson-Mehl and Apollonius Model
point patterns. In particular, we mention
The nearest-neighbor rule (or equiva- Tuominen [1949] and Snyder [1962] (ap-
lently, the crystal growth model men- plications to urban planning), Gambini
tioned earlier) forces the Voronoi regions [19661 and Boots [1979] (market areas;
to be convex polyhedra. Allowing the generalized Thiessen polygons are used
crystals to start their growth at different there, in particular, the Johnson-Mehl
times gives rise to hyperbolically shaped and Apollonius models), l~ollison [1977]
regions. The resulting Johnson-Mehl (ecological contact models), McLain
model was proposed by Johnson and Mehl [19761 (spatial interpolation), Arnold and
[1939] as a more realistic model of struc- Milne [1984] (cartography), and Okabe et
ture for minerals. Figure 6 gives an il- al. [1988] (facility location). For surveys
lustration; crystals are augmented with of Thiessen polygons from a geographical
their date of birth. and economic point of view, see Boots
crystals growing simultaneously but [1986], Eiselt and Pederzoli [1986], and
at different rates give rise to spherically Sibson [1979].
shaped regions forming the Apollonius Suzuki and Iri [1986] report on the
model. This structure can also be ob- importance of recovering the sites from a
served as cell structures of plants or in given subdivision of a geographical area.
foams made out of soap bubbles [Matzke This inverse process of constructing
and Nestler 1946; Smith 1954; Williams Thiessen polygons arises in the optimal
1968]. It further appears as covering ar- outline of school districts or of voting
eas of plants and as areas of best- precincts. In geographical variation
received transmitters [Sakamoto and analysis, connectivity graphs for sites are
Takagi 1988]. Weaire and Rivier [1984] a valuable tool [Matula and Sokal 1980].
give a comprehensive survey and various Among them are the minimum spanning
statistical data. Interestingly, the equi- tree or Prim shortest connection network
librium state of a spider web constitutes (Figure 8), the ~elaunay triangulation,
a generalized, although still polygonal, and the Gabriel graph. These three

e
\ \ o
\ ‘\
\ \
\
\
e
53
—-
--\
-%
● \
Lo \
\
\
\
\
/
/
/
/’ \/
‘i
\
\
Figure 6. Johnson-Mehl model.
concepts are intimately related to contours obtained in this way with per-
Thiessen polygons as will be discussed ceptual boundaries studied by Gestalt
later. psychologists. As pointed out by Blum
t1973], curvature properties of a given
1. 1.5 Blum’s (Medial Axis) Transform
contour correspond to topological proper-
Yet another name for the Voronoi dia- ties of its transform (compare Figure 9).
gram is popular in the natural sciences: The smooth pieces of the contour (solid)
Blum’s transform of a set of sites. Con- should be viewed as sites of generalized
cerned with biological shape and visual shape. By definition, each point of the
science, Blum [1967] used it for modeling transform (dashed) is equidistant from at
new descriptors of shape. In general, the least two sites. In this context, the trans-
main problem in pattern recognition is to form often is called the medial axis or
extract characterizing elements from a skeleton of the contour. The medial axis
given (site) pattern. When no geometri- of digitized contours was exploited in
cal model of the pattern is available, its image processing by Philbrick [19681,
structural description may be based on Montanari [1968], and Lantuejoul and
the notion of neighborhood of a site and Maisonneuve [1984].
thus, in particular, on Blum’s transform. Although the foregoing list of applica-
Fairfield [1979] successfully compared tions of Voronoi diagrams in the natural

Voronoi Diagrams “ 353
2a e
26
88 L3
e
e
55
Figure 7. Spider web
1’
Figure 8. Minimum spanning tree for the nine
state capitals of Austria.
Figure 9. Contour and its medial axis [Philbrick

sciences is not complete, we refrain from 19681.
further details and turn now to their role
in the mathematical environment.
exploited by Dirichlet [18501 to establish,

1.2 Mathematician’s Viewpoint
among other results, a simple proof of
The very earliest motivation for the study the unique reducibility of quadratic
of Voronoi diagrams stems from the the- forms. A careful generalization to higher
ory of quadratic forms. Gauss [18401 dimensions was provided by Voronoi
observed that quadratic forms have an [1908]. To honor the pioneering work of
interpretation in terms of Voronoi these mathematicians, the construct has
diagrams for sites that are parallelohe- been referred to as the Dirichlet tessell-
drally ordered in space. His idea was ation or the Voronoi diagram, and indeed

these terms are the most popular ones in

the vast body of literature.
The following review provides basic
mathematical properties of Voronoi dia-
grams and their generalizations (many of
them will also be crucial in the under-
standing of their algorithmic properties
and applications), as well as points out
certain features of Voronoi diaqams that
have potential applications but may be
not familiar to computational geometry
researchers.
1.2.1 Regularly Placed Sites
Mathematical interest in Voronoi dia-

grams can be divided according to
whether regularly or irregularly placed
sites are involved. A main area where Figure 10. Plesiohedron with 18 facets [Griinbaum
Voronoi diagrams for regularly placed and Shephard 1980].
sites have been used for a long time is
geometrical crystallography.
Let R d denote the d-dimensional
Cartesian s ace. A tiling of Rd is a cov- a stereohedron in R d is finite; for in-
ering of R Y by closed sets whose interi- stance, 390 is an upper bound for d = 3.
ors are pairwise disjoint. Tilings of R d So, by Tarski’s [19511 decidability theo-
by convex polyhedra with the property rem, all spatial stereohedra can be “ef-
that the group of motions mapping the fectively” determined. In spite of this
tiling onto itself is a crystallographic fact, their enumeration is still incom-
group [Bieberbach 1912] are of particular plete. Also, an enumeration method for
interest. Their polyhedra are necessarily plesiohedra mentioned in Delaunay
congruent and were called stereohedra [19631 does not yield a practical algo-
by Federoff [1885] and fundamental rithm. Among others, Nowacki [1976],
domains by Schonflies [1891]. One of Koch [1973], and Engel [1981] discovered
the central questions of geometrical crys- spatial plesiohedra with 18, 23, and 38
tallography has been to enumerate facets, respectively. Figure 10 depicts a
all stereohedra and to classify them plesiohedron with many facets. It seems
according to their crystallographic surprising that its congruent copies are
groups. capable of filling three-space completely.
Most progress was made for the sub- The methods used to obtain plesio-
class of plesiohedra. They can be inter- hedra with large numbers of facets are
preted as Voronoi regions and were called probably as interesting as the numbers
special fundamental domains by themselves. After carefully choosing a
Schonflies [1891] and domains of action set of regularly placed sites (dependent of
by Niggli [1927]. Laves [1930] and several variable parameters), certain
Delaunay et al. [1978] enumerated all parts of its Voronoi diagram are con-
plesiohedra in R2. Applying a refined structed via computer [Engel 1981]. For
classification scheme, Delaunay [1932] excellent sources on tilings, refer to
distinguished among 24 types of plesio- Grunbaum and Shephard [1980, 1987],
hedra in R3. A complete list of It is worthwhile to mention that
planar stereohedra was given by Voronoi diagrams for regularly placed
Grunbaum and Shephard [1987]. sites also find other applications in
Delaunay [1932] showed that the num- mathematics. They apply to numerical
ber of facets (faces of dimension d – 1) of integration [Babenko 1977], to packing

and covering problems for congruent

/. ,? ‘I,
(–)
spheres [Rogers 1964], and to statistical lfl’
investigations of lattice systems [Besag ,’1”
1974; Mollison 1977; Conway and Sloane 111
l,\
1982].
4
,,11
,,1
,,1
,,1 /’
I.
1.2.2 Irregularly Placed Sites \ : ‘ .’”
Let us now turn our attention to Voronoi
diagrams arising from sets of irregularly
placed sites. A tiling of R d by polyhedra
is called a cell complex in R d if the tiling
is facet to facet, that is, each facet of a
polyhedron is also a facet of some other
polyhedron in that tiling.2 Every tiling Figure 11. Voronoi diagrams are related to convex
hulls
of Rd by plesiohedra is a cell complex.
The Voronoi diagram for irregularly
placed sites is still a cell complex
although its regions are no longer graph. The size of higher dimensional
congruent polyhedra. diagrams has only recently been ana-
This fact gives rise to various ques- lyzed completely. Klee [1980] and Brown
tions. Motivated by the problem of find- [19801 showed that Voronoi diagrams in
ing densest sphere packings, Rogers R d are equivalent to certain convex
[19641 posed the following extremal prob- polyhedral surfaces in R ‘f 1 in a strong
lem: How big is the smallest possible sense. Hence, known results on the size
Voronoi region with t facets and defined of polyhedra carry over, in particular,
by sites with minimum distance two? Es- the so-called upper and lower bound the-
timates of this and related quantities orems [Brondsted 19831. Exploiting this
were derived by Muder [1988a, 1988b] for relationship, exact bounds on the num-
planar regions. bers of individual faces of d-dimensional
Much work on general Voronoi dia- Voronoi diagrams were derived by Seidel
grams is concerned with their combinato- [1982] and by Paschinger [1982]. Fig-
rial properties and, in particular, with ure 11 illustrates Brown’s original trans-
their size, that is, their numbers of faces form that relates the Voronoi diagram
of various dimensions as a function of the in R2 to a convex hull in R3 via
number n of sites considered. The size of stereographical projection.
a Voronoi diagram is an important quan-
tity since it relates the amount of space 1.2.3 Generalized Voronoi Diagrams
needed to store this structure to the in- Interestingly, not every convex polyhe-
put size. Answering Crum’s problem, dral surface in R‘+ 1 is related to a
Dewdney and Vranch [1977] exhibit an Voronoi diagam in R ~. For a more gen-
arbitrarily large set of Voronoi polyhedra eral class, a one-to-one correspondence
in R 3 each pair of which shares a facet. can be established. Paschinger [1982] and
The example they give shows that the Aurenhammer [1987a] showed that
size of a Voronoi diagram in R3 is 0( n2); power o!iagrams are equivalent to the
this was also observed by Preparata boundary projection of convex polyhedral
[1977]. As already mentioned, a Voronoi
surfaces. 3 This generalization assigns a
diagram in the plane has size 0(n) since particular weight w to each of the given
its edges and vertices form a planar sites p and replaces the euclidean dis-
tance 6(x, p) between a point x and p by
2 In fact, a tiling if facet to facet if and only if it is

face to face for faces of any dimensions less than d 3 Such projections are commonly called Schlegel
[Gruber and Ryskov 19871. diagrams in discrete geometry [Griinbaum 19671,
ACM Computing Surveys, Vol. 23, No 3, September 1991

Figure 12. Apollomus model [Aurenhammer and Edelsbrunner 1984].
62( x, p) – w. Intuitively speaking, w ex-

presses the capability of p to influence
its neighborhood. In particular, the pair
(p, w) may be interpreted as a sphere
with center p and radius ~ if w > 0.
As an important property, the separator
defined by two weighted sites in R2 is
still a straight line: the power line of two
circles. Figure 4 shows a power
diagram in the plane. Power diagrams
already appeared in Dirichlet [18501—
therefore, often being called generalized
Dirich let cell complexes— and continued
to be objects of interest, especially in
sphere packing [Rogers 1964], illumin-
ating spheres [Linhart 19811, and Figure 13. Planar order-2 Voronoi diagram
the geometry of numbers [Gruber and [Shames19781.

Lekkerkerker 1988].
The concept of weighting the sites gives
rise to several other useful types of dia- n – 1, the structure is often referred to
grams. Weighting the euclidean distance as the furthest site Voronoi diagram since
by a multiplicative constant yields the now the region of a site contains all points
Apollonius model, investigated by in space furthest from it. Figure 13 shows
Aurenhammer and Edelsbrunner [1984]. an order-2 Voronoi diagram for 8 sites in
The separator of two sites in the plane the plane. Note that only 15 of the 28
describes their Apollonius circle. From pairs of sites define a region and that a
Figure 12 it can be seen that regions separator may yield more than one edge.
may be disconnected and may partition (An example of a furthest site Voronoi
the plane into G( n2) connected compo- diagram is given in Figure 24.)
nents. Additiue weights give rise to the Early treatments of order-k diagrams
Johnson-Mehl model (compare Figure 6). are found in Miles [1970] and in Shames
For many additional types, see Ash and and Hoey [1975]. Lee [1982al succeeded
Bolker [1986]. in proving exact upper bounds on the
At this point, let us modify the concept maximum numbers of regions, edges, and
of Voronoi diagram in another way. The vertices of a planar order-k Voronoi
order-h Voronoi diagram of n sites is a diagram. Asymptotically these num-
partition of R d into regions such that bers are @( k( n – k)). The fascinating
any point within a fixed region has the relationship between Voronoi dia-
same k closest sites. Its regions are con- grams, weighted diagrams, order-k dia-
vex polyhedra that form a cell complex grams, and other geometric objects was
in R‘. For k = 1, the classical Voronoi worked out by Edelsbrunner et al.
diagram is obtained. In the case of k = [19861, Aurenhammer [1987 al, and

Aurenhammer and Imai [1988]. Within sphere and on the torus, respectively, are
this context, the concepts of power dia- closely related to their equivalents in the
gram and hyperplane arrangement play euclidean space of the same dimension.
a central role. Section 3.1 is devoted to a Diagrams on three-dimensional poly-
detailed description of this material. In hedral surfaces and on the three-
particular, various results on the size and dimensional cone have been treated by
on the computational complexity of con- Mount [1985] and by Dehne and Klein
structing such diagrams are obtained. [19871, respectively.
However, still little is known on the size
of the order-k Voronoi diagram in higher 1.2.4 Delaunay Triangulations
dimensions; consult [Edelsbrunner and
Seidel 1986]. Only for the furthest-site Hand in hand with the investigation of
diagram exact upper bounds are avail- Voronoi diagrams goes the investigation
able [Seidel 1987]. The number of re- of related constructs. Among them, the
gions of an order-k Voronoi (or power) Delaunay triangulation is most promi-
diagram is closely related to the num- nent. It contains a (straight-line) edge
ber of k-sets of a finite point set. Deter- connecting two sites in the plane if and
mining these numbers belongs to the only if their Voronoi regions share a
main open problems in combinatorial common edge. The structure was intro-
geometry. duced by Voronoi [1908] for sites that
Once the usefulness of known types of form a lattice and was extended by
Voronoi diagrams had been realized, fur- Delaunay [1934] to irregularly placed
ther generalizations were attempted. sites by means of the empty-circle me-
Motivation mostly came from applica- thod: Consider all triangles formed by
tions in computational geometry. In or- the sites such that the circumcircle of
der to meet practical needs, general each triangle is empty of other sites.
distance functions like the LP-metrics The set of edges of these triangles
[Lee 1980a] or convex metrics [Chew gives the Delaunay triangulation of
and Drysdale 1985] were taken to define the sites.
a Voronoi diagram. Also, the shape of The planar Voronoi diagram and the
sites was varied while using a canonical Delaunay triangulation are duals in a
extension of the euclidean distance func- graphtheoretical sense, Voronoi vertices
tion. Line segments [Kirkpatrick 1979], correspond to Delaunay triangles,
curve segments [Yap 1987], and disks or Voronoi regions correspond to sites, and
general convex sites [Leven and Sharir edges of both types correspond by defini-
1986, 1987] have been considered. If the tion. From Figure 14 it can be seen that
segments form a closed curve, their dia- Delaunay edges (solid) are orthogonal to
gram is just the medial axis of a contour their corresponding Voronoi edges
(compare Figure 9). Disks can be viewed (dashed)–but do not necessarily inter-
as centers weighted additively by their sect them— and that the boundary of
radii, thus giving rise to the Johnson- the convex hull of the sites consists of
Mehl model. This list is not meant to be Delaunay edges. The duality immedi-
exhaustive; we shall be concerned with ately implies upper bounds of 3 n – 6 and
mathematical and computational proper- of 2 n – 4 on the number of Delaunay
ties of varous generalized Voronoi edges and triangles, respectively (com-
diagrams in later sections. pare Introduction). The Delaunay tri-
Let us mention one more way of gener- angulation and its duality to Voronoi
alization: changing the underlying space. diagrams generalize to higher dimen-
Ehrlich and Im Hof [1979] investigated sions in an obvious way. By results
the behavior of Voronoi regions in by Dewdney and Vranch [1977], the
Riemann manifolds. Brown [1980], Delaunay triangulation in R3 may
Paschinger [1982], and Yap [1987] ob- already be the complete graph on
n
served that Voronoi diagrams on the n sites, thus having edges. For a
()2

angulations without “extreme” an-

\ gles are desirable in finite element
/
‘i\ // and interpolation methods.
e
/’ Lawson [1972] gave a counterexample
I ‘. /
1’ /
/ to Shames and Hoey’s [1975] conjecture
\ / that the Delaunay triangulation has
‘Y’ minimum total edge length. It does not
\
_- /’ \
, A even approximate the shortest triangula-
--! -.
\ ~y/ ‘.\ tion [Manacher and Zobrist 1979], and in
----
x.
\’~ ‘. fact it may be as long as any triangula-
Y’ \ . .
tion [Kirkpatrick 1980]. It is, however,
/’ \
/ \ close to optimal on the average [Lingas
/ \ 1986a]. A different criterion of optimal-
/ \
ity is mentioned in McLain [1976]. For
each triangle, all its points should be at
Figure 14. Voronoi diagram and Delaunay trian-
gulation are duals,
least as close to one of its defining sites
as to any other site. This property is not
shared by the Delaunay triangulation, as
is claimed there.
catalog of properties of higher dimen- As an important fact, the Delaunay
sional Delaunay triangulations arising triangulation is a supergraph of several
from site lattices see Gruber and well-known and widely used graphs
Lekkerkerker [19881. spanned by a set of sites in the plane.
Several interesting properties are Among them are the minimum spanning
known for the Delaunay triangulation tree (or Prim shortest connection net-
in the plane. It was first observed by work ) introduced by Kruskal [1956] and
Sibson [1977] that this triangulation Prim [1957], the Gabriel graph intro-
is locally equiangular. This property duced by Gabriel and Sokal [1969], and
holds when, for any two triangles whose the relative neigh borhood graph intro-
union is a convex quadrilateral, the reduced by Toussaint [19801. Section 2.4
placement of their common edge by the gives definitions of these graphs and
alternative diagonal does not increase the describes their interrelations.
minimum of the six interior angles con- The Delaunay triangulation can be ex-
cerned. Actually, the Delaunay triangu- ploited to find certain linear combina-
lation is the only one with this property tions among its defining sites. Each site
that particularly shows its uniqueness. not on the convex hull can be repre-
For sites being in general position (the sented as a weighted mass center of the
Delaunay triangulation may contain sites adjacent in the triangulation; see
more-sided faces if four sites are cocircu- Sibson [1980] who mentions applications
lar), Edelsbrunner [19871 showed that lo- to surface smoothing. Aurenhammer
cal equiangularity is equivalent to global [1988b] generalized this result to power
equiangularity. The triangles define the diagrams and their order-k modifica-
lexicographically largest list of sorted tions. In particular, Gale transforms of
angles. A similar result holds if only the the sites may be derived in this way
smallest angle of each triangle is consid- [Aurenhammer 1990bl. Gale transforms
ered [Lawson 1977]. Note that the Delau- are a versatile tool in the investigation of
nay triangulation thus maximizes the high-dimensional convex polyhedra
minimum angle over all triangulations [Grunbaum 19671.
of a given set of sites. On the other hand, The generalized Delaunay triangula-
a simple example shows that the maxi- tion obtained from power diagrams gains
mum angle is not minimized. It is inter- in importance from the following recog-
esting to note that, by the empty circle nition problem: Given some cell complex,
property, any triangulation without can it be interpreted as a Voronoi dia-
obtuse angles must be Delaunay. Tri- gram? A cell complex in R d may be

viewed as a power diagram if and only if able paper, Meijering [1953] derived
the cell complex admits a certain dual means for the volume, the total boundary
construct, called the reciprocal figure by area, and the total edge length of the
Maxwell [1864], Crapo [1979], Whiteley regions of a Voronoi diagram in R 3 aris-
[1979], and Ash and Bolker [1986]. The ing from a Poisson field of sites, as well
reciprocal figure is completely character- as the average number of vertices, edges,
ized by the properties of the Delaunay and facets. Several of these quantities
triangulation for power diagrams. Using are also given for the Johnson-Mehl
reciprocal figures as a criterion, model. Further progress was made by
Aurenhammer [1987b] showed that sev- Gilbert [19611 who determined the vari-
eral well-known types of cell com- ances for the volumes of such Voronoi
plexes are power diagrams. Moreover, regions. He also found some variances
Voronoi diagrams can be recognized associated with plane or line section
and their defining sites can be re- through regions. Based on experimental
stored in an efficient way using recip- results, Kiang [19661 gave a formula for
rocal figures [Aurenhammer 1987c]. For the random size distribution of Voronoi
a nice survey on reciprocal figures, regions in R 1, along with a conjecture for
see Ash et al. [1988]. Essentially distinct its generalization to R2 and R3. The
criteria for cell complexes to be classical conjecture is, however, incorrect accord-
or to be multiplicatively or additively ing to results in Gilbert [1961]. Figure 15
weighted, Voronoi diagrams were pro- shows his obtained density functions in
posed by Ash and Bolker [1985, 1986]. R1 (solid), in R2 (dashed), and in R3
Blum [1967] first suggested that gener- (dotted). Observed volume over mean
ally shaped sites could be reconstructed volume is scaled.
from the shape of the regions they de- Miles [1970] made a thorough study of
fine. Calabi and Hartnett [1968] elabo- Voronoi diagrams induced by a planar
rated on this question in some detail. homogeneous site process. Expectations of
Recognition problems of this kind find the edge number, edge length, perimeter,
applications, aside from geography and and area of the regions are given.
economics, in statics and in computer Further results concern the Delaunay
science. triangulation and the order-k family of
This list of geometric and combina- diagrams. Relevant experimental data
torial properties of Delaunay triangu- can be found in Crain [19781. His ob-
lations is not complete. We shall served frequencies of edges per polygon
see various others while discussing are of particular interest since no theo-
applications in computational geometry. retical results are presently available.
An important result by Dwyer [1989]
shows that the expected number of ver-
1.2.5 Stochastic Properties
tices of the classical (closest or furthest
So far, we have surveyed a good deal of site) Voronoi diagram in d dimensions is
research on geometric and combinatorial. only 0(n) if the n sites are uniformly
aspects of Voronoi diagrams. Another drawn in a hyperball. At this place, we
significant stream of investigations con- mention only marginally that several re -
cerns the determination of statistical data suits have been obtained that concern
about the diagrams obtained from ran- the behavior of Voronoi diagrams for sites
dom distributions of sites. One of the being introduced in random order rather
earliest and strongest motivations for than being drawn by some probability
studying stochastic properties of Voronoi distribution. We shall see such results in
diagrams stems from their practical rele- Section 1.3.
vance to physical and chemical processes, Space constraints preclude mentioning
especially in metallurgy and crystallog- all related work. To name a few authors,
raphy [Johnson and Mehl 19391. Accord- we refer to Newman et al. [19831 (num-
ingly, most efforts concentrated on sites ber of nearest neighbors in d di-
distributed in R2 and in R3. In his valu- mensions), Weaire and Rivier [19841

360 . Franz Aurenhammer
Figure 15. Empirical density function of the volume of regions [Kiang 19661.
(generalized diagrams), Besag [19741, intuitive construction rules were pro-

Mollison [19771, Cruz Orive [19791, posed. The earliest diagrams were drawn
Conway and Sloane [1982] (site with pencil and ruler; see Horton [1917]
lattices; applications to spread of or Kopec [19631 who mentions problems
epidemics and to coding theory are of ambiguity if many sites lie on a com-
given), and Aurenhammer et al. mon circle. Probably the most obvious
[19911 (probabilistic distance functions). approach is to delineate the diagram in
The reader interested in geometrical the plane region by region, by singling
probability in general and stochastic out those separators that contribute to
properties of cell complexes in particular edges of the current region. A provi-
may consult the expository papers by sional and admittedly inefficient version
Moran [1966, 19691, Little [19741, and of such an algorithm was described by
Baddeley [1977]. Rhynsburger [19731. Other early algo-
rithms build up the diagram vertex by
1.3 Computer Scientist’s Viewpoint vertex [Brassel and Reif 1979] or by in-
cremental insertion of sites (i. e., of their
We have documented the remarkable role
regions) [Green and Sibson 19771.
that Voronoi dia~ams play in the math-
Figure 16 illustrates the insertion of a
ematical and applied natural sciences.
site p that involves two tasks. First, we
Yet for a long time their practical useful-
need to find the current region in which
ness suffered from the absence of reason-
p falls. Let q be the site defining this
ably simple and efficient methods for
region; the separator of p and q then
their computation. This section reviews
will contribute an edge, e, to p’s region.
methods for the computer construction
Second, the boundary of p’s region (bold)
and representation of Voronoi diagrams.
is created edge by edge, starting with e.
Algorithmic applications of Voronoi
During this process, the parts of the old
dia~ams and of related structures are
diagram closest to p (dashed) are
discussed in Section 2.
traversed and deleted. These parts are
specified by lying on p’s side of
1.3.1Early Algorithms
the separators of p and its new neigh-
Since Voronoi diagrams have been used bors. If appropriately implemented, the
for decades by natural scientists, many second task requires time proportional to

diagram in R d that grows exponentially

with d. A three-dimensional implemen-
tation of Watson’s algorithm is discussed
\ in Field [1986]. Avis and Bhattacharya
\
\ [1983] propose an O(n rdlzl ‘1) time algo-
\ rithm for determining all vertices of the
,“\
e d-dimensional diagram. They also out -
line a linear programming method for
calculating the d-dimensional Delaunay
e
triangulation but give no concrete
complexity analysis.
Only recently have optimal or near-
optimal algorithms for constructing
Voronoi diagrams been devised. The rea-
son is that only in the last few years
have powerful algorithmic techniques
Figure 16. Inserting a Voronoi region. been fully developed and exploited for
computational geometry purposes.
the number of edges deleted. This num- 1.3.2 Speeding Up Insertion

ber is 0(i) in the worst case (i denotes
The process of building a Voronoi dia-
the number of sites inserted so far) since
gram in the plane by incremental inser-
we have to delete a planar graph with at
tion of sites stands out by its simplicity.
most i regions. The first task requires
Originally having an 0( n ~) expected
0(i) time in the worst case, but a simple
performance, the Green-Sibson algorithm
heuristic will achieve 0(V) expected
may be polished up to run faster for sev-
time. Experimental results showed that
eral distributions of the sites. By intro-
this method is quite efficient and thus is
ducing suitable orderings of the sites, the
applicable to rather large sets of sites. As
expected time for finding the region the
for the other algorithms mentioned
next site falls in and for integrating the
above, however, the worst-case behavior
new region can be lowered to O(1) [Ohya
is 0( nz) for n given sites.
et al. 1984a, 1984b].4 This gives an ex-
Several intuitive methods are also
pected runtime of 0(n). Clearly this is
available for the Voronoi diagram in R 3.
the best we can hope for because the size
They compute the induced cell complex
of the diagram is a trivial lower bound
facet by facet [Brostow et al. 1978], ver-
for the time needed to compute it.
tex by vertex [Finney 1979], or via its An alternative approach to speeding
dual, the Delaunay triangulation
up insertion is randomization. Based on
[Tanemura et al. 19831. Potentially, they a general result by Clarkson and Shor
require 0( n4) time in the worst case,
[19881, Mehlhorn et al. [19901 showed that
although they may perform quite well for inserting the sites in random order yields
various distributions of sites. For d >4 an 0( n log n)-time performance with
dimensions, insertion strategies working
high probability. This complexity is inde-
directly [Bowyer 19811 or based on the pendent of the distribution of the sites;
Delaunay triangulation [Watson 1981] expectation is taken over all possible
have been used to construct a suitable
permutations of the sites. The algorithm
combinatorial representation of the extends to the class called abstract
Voronoi diagram. Their time complexity
is O(nl+ll~) and 0(n2– I[d), respectively,
provided the sites are “well distributed. ”
This should be contrasted, however, 4 Recall in this context that the average number of
against the maximal size of a Voronoi edges of a region is less than six.

362 * Franz Aurenhammer
Voronoi diagrams by Klein [1989] with-

out increase of runtime. This general
“{
concept includes power diagrams and di-
agrams defined by line segments or by e !-
/ :.-
● “.,,. .,.,
LP-metrics. Guibas et al. [1990] propose ----- ----- I,
.
an even more practical version of ran- -.. .:.’---,

‘, ● >. :0
domized incremental construction of the ‘. // .,.
‘y .’ .,..-. .,
/)--:. -{,‘1* .:
classical Voronoi diagram. “
As with other geometrical algorithms, ! “ ‘“ .,
the problem of numerical errors arises in /
/’ .’
the construction of Voronoi diagrams. For /

,{ ,,,
example, sites nearly lying on a common
circle define vertices that tend to ap-
proach arbitrarily close. A proposal for
making the strategy of insertion robust Figure 17. Merging two Voronoi diagrams
against numerical errors is outlined in [Shames 19781.
Sugihara and Iri [1988].
The process of inserting Voronoi re-
gions extends nicely to R3. Regions are
left sites (dashed) with the diagram for
convex polyhedra that can be constructed
facet by facet by intersecting existing right sites (dotted).
Section 3.2 is devoted to a detailed
regions with separators that are planes
in this case. A region cannot have more description of the method and its gener-
than n – 1 facets (one for each different alizations. In particular, two diagrams
site) and thus by Euler’s relation, has can be merged in 0(n) time, which im-
0(n) edges and vertices. Therefore, one plies an 0( n log n) time algorithm if the
needs 0( n + t) time per region if t facets, divide step is carried out in a balanced
edges, or vertices are deleted during
way. This is worst-case optimal since any
its insertion; see Aurenhammer and
algorithm that constructs an explicit de-
Edelsbrunner [1984] who treat insertions
scription of a Voronoi diagram may be
in a more general cell complex in R3. used to sort: Interpret an input sequence
Since each component deleted has to be of n real numbers as a set of sites on the
constructed first, an 0( n2 )-time algo- x-axis and construct their Voronoi dia-
rithm results. This is worst-case optimal gram. Regions of sites for consecutive
since a Voronoi diagram in R3 may have numbers will share an edge, so the sorted
a size of 61(n2). sequence can be obtained in 0(n) time
by scanning through the regions.5
Although theoretically fast and ele-
1.3.3 Divide and Conquer gant, the divide-and-conquer approach
has certain disadvantages. Implementa-
A widely used method to design fast al-
tion details are somewhat complicated,
gorithms is divide and conquer. Shames
and numerical errors are likely by
and Hoey [19751 observed that this
method applies well to problems in com-
putational geometry and, in particular,
to the construction of a Voronoi diagram
in the plane: The given set of n sites is 5 Voronoi diagrams do not help for sorting general
sets of sites. Seidel [19851 proved that sorting n
divided into two subsets by a vertical sites in the plane with respect to their x-coordi -
line. The diagrams for these subsets are nates takes Q( n log n) worst-case time even when
computed recursively and are then their Voronoi diagram is part of the input. It is
“merged” in the conquer step to form the unclear whether sorting helps for Voronoi dia-
grams Fortune [1988], however, pointed out that
total diagram. Figure 17 shows the chain
presorting the sites in two different directions low-
of edges (bold) to be constructed during ers the additional time for constructing the
the process of merging the diagram for diagram in the L1-metric to 0( n log log n),

construction: Dividing the plane into [19771 that requires O(n log n) time. The
narrow slabs forces the vertices defined convex hull is dual to the Voronoi dia-
by sites within a slab to approach infin- gram of the given sites in a geometrical
ity. Moreover, the expected behavior is sense. Hence 0(n) additional time suf-
still 0( n log n). An 0( n)-expected-time fices for deriving the latter. This elegant
algorithm that applies divide and con- approach matches the optimal 0( n log n)
quer to a certain subset of sites was time bound, although it suffers from the
outlined by Bentley et al. [1980]. problem of processing three-dimensional
Divide and conquer is the basis of a objects.G
large class of algorithms for computing An important feature of the embedding
generalized Voronoi diagrams in the method is its easy generalization to
plane. Let us briefly list some of them in higher dimensions. Methods for deter-
order to give credit to the authors who mining higher dimensional convex hulls
considered them first. Shames and Hoey are well established [Seidel 1981, 19861,
[19751 reported that their algorithm also hence efficient worst-case algorithms for
applies to the furthest site Voronoi dia- computing the d-dimensional Voronoi di-
gram in the plane. Hwang [1979], Lee agram become available. Their runtime,
and Wong [1980], and Lee [1980] consid- however, increases exponentially with d
ered point sites under the LP-metrics, according to the maximal size of a dia-
whereas even more general distance gram. This should be contrasted with a
functions were treated by Widmeier et recent result by Dwyer [1989], showing
al. [1987], Chew and Drysdale [1985], and that 0(n) expected time suffices for com-
Klein and Wood [19881. Imai et al. [1985] puting the Voronoi diagram in constant
treated sites under the Laguerre dis- dimensions d for uniformly distributed
tance; that is, their power diagram. sites.
Concerning the divide-and-conquer con- Brown’s idea was further developed by
struction of Voronoi diagrams for sites Edelsbrunner et al. [1986] for construct-
more general than points, we refer to ing the order-k Voronoi diagram family
Kirkpatrick [1979] (line segments), Lee and by Aurenhammer [1987al for con-
and Drysdale [1981] (line segments or structing the power diagram and its
nonintersecting discs), Sharir [1985] order-k family in general dimensions.
(arbitrary disks), Leven and Sharir [19871 See also Edelsbrurmer [19861, Aggarwal
(planar convex bodies), Yap [1987] (curve et al. [1989a], and Aurenhammer [1990al
segments), and Lee [1982b] (medial axis where order-k diagrams are obtained in
of a simple polygon). It should be noted different ways from their higher dimen-
that most of the algorithms just mensional embeddings. We refrain from any
tioned are worst-case optimal. details here and refer to Section 3.1 for a
comprehensive discussion of this mate-
rial. For different approaches to the com-
1.3.4 Higher Dimensional Embedding
putation of order-k diagrams in the plane
Transforming geometrical problems into consult Lee [1982al and Chazelle and
more easily understood and solved ones Edelsbrunner [19871. Clarkson [19871
plays an important role in computational speeds up a combined version of these
geometry (e.g., see [Edelsbrunner 19871). approaches using random sampling,
Brown [1979] first perceived the possibil- achieving O(knl”) expected time and
ity of transforming Voronoi diagrams in
R2 into convex hulls in R3: The sites are
mapped, via stereographical projection, 6 In fact, since Preparata and Hong construct con-
into points lying on a sphere (compare vex hulls by divide and conquer, this is just a
Figure 11). This takes 0(n) time. three-dimensional translation of a divide-and-
conquer construction. By Brown’s result, simple
The convex hull of the resulting three- convex hull algorithms [Clarkson and Shor’s 1988]
dimensional point set is computed using lead to simple Voronoi diagram algorithms,
the algorithm by Preparata and Hong however.

364 ● Franz Aurenham mer
space (for any c > O) that is nearly

optimal.
Embedding in higher dimensions yields
algorithms for various different types
of diagrams. See Aurenhammer and
Edelsbrunner [1984] for the multipli-
catively weighted Voronoi diagram,
Aggarwal et al. [1989a] for the medial
axis of a convex polygon, Edelsbrunner
et al. [19891 for cluster Voronoi dia-
grams, and Aurenhammer and Stockl
[1988] for the peeper’s Voronoi
diagram.
Figure 18. Maintaining the intersection with a

1.3.5 Plane-Sweep Construction line.
Another powerful technique in computa-

tional ge~metry is the pjane-sweep ‘tech-
nique [Preparata and Shames 19851. In conquer algorithms [Lee and Lin 1986;
contrast to the embedding method, this Lingas 1986b; Chew 1989a; and Wang
technique decreases the dimension of the and Schubert 19871. As was reported by
problem. Intuitively speaking, the static Rosenberger [19881, plane sweep also
problem of computing a Voronoi diagram performs well for planar order-k dia-
in the plane is reduced to the dynamic grams even when the sites are weighted
problem of maintaining the cross section additively.
of the diagram with a straight line. The
algorithm simulates sweeping a line
1.3.6 Delaunay Triangulation Algorithms
across the plane from below. At any point
in time, the portion of the diagram below Since the Voronoi diam-am and the
the sweep line is complete (Figure 18). Delaunay triangulation ‘are duals, the
Fortune [1985, 1987] first observed that combinatorial structure of either struc-
updates on the sweep line can be imple- ture is com~letelv determined from its
mented to cost O(log n) time if a certain dual. Conse~uentfy, the Delaunay trian-
continuous deformation of the diagram is gulation of n sites in the plane can be
treated. From this deformation, the orig- obtained in O(n) time after an
inal diagram can be constructed in 0(n) 0( n log n)-time precomputation of the
time. The method, its application, and its Voronoi diagram. Several practical ap-
modification for generalized Voronoi plications, however, solely exploit combi-
diagrams are described in detail in natorial properties of the Delaunay
Section 3.3. triangulation. This has led to the
The plane-sweep approach to Voronoi investigation of methods for constructing
dia~ams combines simplicity and effi- this structure directly, thus avoiding
ciency. It achieves the optimal 0( n log n) the need to calculate and store the
time and space bounds for the clas-
0(n) coordinates of Voronoi vertices.
sical type, the additively weighted type, Lawson [1972] proposed an algorithm
and the dia~am for line segments for constructing locally equiangular tri-
[Fortune 1987]. The method has been angulations by local improvement. Start-
successfully applied by Seidel [1988] ing with an arbitrary triangulation of
to Voronoi diagrams with line segment the sites, edges are “flipped” according
obstacles joining sites (and their con- to the equiangularity criterion (stated in
strained Delaunay triangulations). This Section 1.2 .4) until no more such ex-
type had been previously attacked by changes are required. Sibson [19771 suc-
means of rather involved divide-and- ceeded in proving that this procedure

Voronoi Diagrams w 365
triangulation are maintained to facili-

tate the determination of the triangle
covering the next site to be inserted. This
practical algorithm exhibits a random-
ized running time of 0( n log n) and is
similar in spirit to Clarkson and Shor’s
[1988] convex hull algorithm.
The question arises as to which of the
planar Delaunay algorithms can be ex-
Figure 19. Equiangularity and empty circle prop-
tended to higher dimensions. We have
erty. already seen some early (although widely
used) algorithms at the beginning of this
section. It should be further observed that
the method of higher dimensional em-
actually constructs the Delaunay tri- bedding actually first constructs the
angulation. Lawson’s original method Delaunay triangulation (via a convex
was somewhat speeded up by Lee and hull), then derives the Voronoi dia-
Schachter [1980] who proceed incremen- gram via dualization. Recently, Rajan
tally and apply the edge flipping proce- [19911 proved a property of d-
dure after each insertion of a new site. dimensional Delaunay triangulations re-
McLain [19761 uses the empty circle lated to equiangularity. It implies an
property to construct the Delaunay trian- incremental algorithm that makes the
gulation by successively adding triangles triangulation locally Delaunay after each
whose circumcircles are empty of sites. insertion of a new site by applying a
In fact, Figure 19 illustrates the equiva- triangle-flipping procedure similar to
lence of the empty circle property and Lawson’s method in the plane. The algo-
the equiangularity criterion. Although rithm is worst-case optimal for odd d.
being simple to implement, all these al- Whether every d-dimensional triangula-
gorithms have an 0( nz)-time behavior in tion can be made Delaunay by local
the worst case. On the other hand, Maus improvement remains open.
[19841 reports that a modification of
McLain’s algorithm based on preparti-
1.3.7 Storage Representation and Dynamization
tioning exhibits an 0(n) expected run-
ning time for uniformly distributed sites, The Voronoi diagram maybe viewed as a
Lee and Schachter [1980] gave an data structure that organizes its defining
0( n log rz)-time Delaunay triangulation sites in a prescribed manner. A host of
algorithm; it uses divide and conquer and applications of this data structure in
is similar to Shames and Hoey’s primal computational geometry will be given in
Voronoi algorithm. A modified imple- Section 2. Out of a number of possibili-
mentation using prepartitioning was ties to store the planar Voronoi diagram
shown by Dwyer [19871 to run in or, equivalently, its dual, the Delaunay
O(n log log n) expected time while re- triangulation, the quad-edge structure
taining the optimal 0( n log n) worst-case proposed by Guibas and Stolfi [19851 is
time; it generalizes to the LP-metrics for particularly practical, Essentially, each
p > 2. Drysdale [19901 modifies the edge is stored as a pair of directed edges.
divide-and-conquer approach to compute, Each directed edge, in turn, stores the
in 0( n log n) time, Delaunay triangula- vertex it originates from and pointers to
tions generated by general convex dis- the previous and to the next (directed)
tance functions. A recent algorithm by edge of the region to its left, The quad-
Guibas et al. [19901 inserts (point) sites edge data structure provides a clean
in random order. Using a structure simi- separation between topological and geo-
lar to the Delaunay tree [Boissonnat and metrical aspects and supports the imple-
Teillaud 19861, earlier versions of the mentation of standard techniques for
ACM Computing Surveys, Vol 23, No. 3, September 1991

366 0 Franz Aurenhammer
computing Voronoi diagrams, such as di- Delaunay triangulation (Figure 21).

vide and conquer or incremental inser- Any triangle T destroyed by insert-
tion. The structure has been generalized ing a new site gets as sons (solid
to higher dimensions by Dobkin and pointers) the new triangles sharing
Laszlo [19891 (facet-edge structure in an edge with T. Remaining triangles ad-
R3) and by Brisson [1989] (cell- jacent to T in such an edge get the re-
tuple structure in R ~). These data struc- spective new triangle as a stepson (dotted
tures represent the incidence and pointers). Guibas et al. [19901 further
ordering information in a cell com- developed this structure. Among other
plex in a simple uniform way. See also results, they showed that the expected
Aurenhammer and Edelsbrunner [1984] number of structural changes of the
and Edelsbrunner et al. [1986] for Delaunay tree is 0(n) if n sites are
different data structures representing inserted in random order into an (ini-
cell complexes. tially empty) Delaunay triangulation.
Like any data structure, the Voronoi A different way of dynamiting Voronoi
diagram can be dynamized, that is, diagrams or Delaunay triangulations is
maintained for a set of sites that varies to allow movement of sites (continuous
over time by insertion or deletion. The updates) rather than their insertion or
case of site insertion is covered by the deletion (discrete updates). The Voronoi
algorithms mentioned before that build diagram of continuously moving sites will
up the diagram on line by introducing change its shape continuously, although
new sites. Integration of a new region only at “critical” points in time will the
into the Voronoi diagram for n sites in combinatorial structure of the diagram
the plane clearly costs 0(n) time in the change. Aside from degenerate cases,
worst case since the region may have n such combinatorial changes always are
edges. Gowda et al. [1983] handle inser- local (Figure 22). An edge between two
tions and deletions of sites in 0(n) time regions (here 2 and 4) collapses to a ver-
by means of the so-called Voronoi-tree tex and reappears between two others
that occupies 0( n log log n) space. This (1 and 3) just in the moment when the
tree records the history of a divide-and- four sites involved get cocircular. In the
conquer construction of the diagram dual environment, this process corre-
(Figure 20). Its leaves hold the sites in sponds to flipping the diagonals of the
lexicographical order. Each inner node is quadrilateral spanned by these four sites,
associated with the diagram that comes which is the union of two Delaunay tri-
from combining its sons’ subdiagrams. angles. This suggests the definition of
The idea originates with Overmars [1981] the Delaunay history of a set of moving
who, however, did not give details of the sites as the chronologically ordered list of
method or its storage requirement. By a edge flips in their Delaunay triangula-
result by Aggarwal et al. [1989a], deletion. Tokuyama [1988] considered the
tion of a given site can be performed in case of two rigidly moving sets of n and
time proportional to the number of edges m sites, respectively. Their Delaunay
of its region. history has length 0( nm) and can be
Using the Delaunay triangulation, a computed and preprocessed in
planar set of sites can be organized into a 0( nm log nm) time in order to retrieve
hierarchical data structure, called the the Delaunay triangulation (and the
Delaunay tree by Boissonnat and Voronoi diagram) in 0( n + m) time at
Teillaud [1986]. This structure is semidy - any given moment. For n sites moving
namic; it allows the insertion of sites and in fixed but individual directions and
supports the retrieval of the triangle in with constant individual velocity, the
which a given point falls in efficient Delaunay history has length 0( n3).
expected and worst-case time. The This result by Imai et al. [1989] should
Delaunay tree reflects the history be contrasted with the intuitive
of an incremental construction of the argument that cocircularities may
():

k~
le
A
*3
2°
L
●
.5
B
6°
97 e’
(-)c
\
\
\
\\ \
\‘)--’) \
.. ___ ‘-----
‘---
“..
~F(6
El El
\
\
Figure 20. Voronoi tree [Gowda et al. 19831.
occur among n sites. Guibas et al. [19911 task since many of the techniques used
showed that the bound remains nearly (incremental insertion or plane sweep,
cubic if the sites are moving continu- for instance) seem inherently sequential,
ously along rather general trajectories. Since it is one of the fundamental struc-
See also Aonuma et al. [1990] for related tures in this area, the Voronoi diagram
results. has been among the first geometrical
structures whose construction has been
parallelized. The underlying model of
1.3.8 Computing Voronoi Diagrams in Parallel
computation mostly has been the CREW
Parallelizing algorithms in computa- PRAM (concurrent-read, exclusive-write
tional geometry usually is a complicated parallel random access machine), where

. ..<.. .
\
‘.
>
Figure 21. Delaunay tree.
I I In her pioneering work, Chow [1980]

showed that convex hulls in three space
---:!$---
“;+...
can be computed in 0(log3 n) time using
/
0(n) processors. By their duality to
planar Voronoi diagrams, the complexity
carries over. Aggarwal et al. [1988] im-
3 proved the Voronoi diagram construction
/ I
to 0(log2 n) parallel time by applying
/ divide and conquer. Their approach criti-
cally depends on restricting the domain
Figure 22. Combinatorial change. that can contain the “merge chain”
(compare Figure 17) and has been im-
proved recently by Cole et al. [1990]
using a tailor-made data structure. An
processors share a common memory optimal parallel construction of Voronoi
that allows concurrent reads but no diagrams is still outstanding, however.
simultaneous writes to the same cell. Goodrich et al. [1989] construct the
As has been mentioned, !2( n log n) is a Voronoi diagram for line segment sites
lower time bound for sequentially con- in 0(log2) time using 0(n) processors.
structing the Voronoi diagram for n sites Their approach to constructing this im-
in the plane. Therefore, O(log n) parallel portant generalization is similar in spirit
time is the best that can be achieved to Yap’s [1987] curve segment algorithm.
when only 0(n) processors are to be used. AS a related question, Schwarzkopf

[1989] addresses the problem of comput- sites. This reduces the problem to finding
ing the digitized image of a planar the region that contains the query point,
Voronoi diagram using certain processor an approach that became powerful once
networks. efficient data structures were developed
for the latter problem: point location in
2. ALGORITHMIC APPLICATIONS
planar straight-line graphs [Kirkpatrick
1983; Edahiro et al. 1984; and
The precomputation of a Voronoi dia- Edelsbrunner et al. 1986]. Actually,
gram is the initial step of various algo- logarithmic query time and linear
rithms in computational geometry. We storage are achieved. This is asymp-
have already mentioned some of them in totically optimal since it matches
the Introduction, In this part of the the information-theoretical lower bound.
survey we demonstrate the broad Shames’ approach in conjunction with
scope of algorithmic applications of Kirkpatrick’s was worked out in a geo-
Voronoi diagrams and of closely related metrically dual setting by Edelsbrunner
structures. and Maurer [1985]. Usage of a structure
related to Boissonnat and Teillaud’s
2.1 Closest-Site Problems
[19861 Delaunay tree even obviates the
need of postprocessing the Voronoi dia-
According to their definition, Voronoi di- gram for point location; Guibas et al,
agrams apply naturally to various prox- [1990] show that post-office queries are
imity problems. Determining closest sites supported in 0(log2 n) expected time by
plays a dominant role in this context this structure. An off-line solution to the
since it appears as a subroutine of many post-office problem (all query points are
geometric algorithms. Applications in- given in advance rather than being spec-
clude clustering and contouring and var- ified on line) was presented by Lee and
ious other problems whose relation to Yang [19791.
proximity may not be so obvious. In principle, higher dimensional vari-
ants of the post-office problem may be
2.1.1 Nearest-Neighbor Queries
solved via the Voronoi diagram, too.
Their practical relevance is evident from
Probably the most popular problem in the equivalent file-searching problem for
this area is the post-office problem. How multiattribute data mentioned in the In-
should a fixed set of n sites in the plane troduction. Although Chazelle [1985]
be preprocessed in order to determine showed that point-location in a Voronoi
quickly the site closest to an arbitrarily diagram in R3 can be performed effi -
chosen point (the query point)? The post- ciently, the approach suffers from a con-
office problem is motivated in the siderable storage overhead of @( n2 ) that
Introduction as a two-dimensional comes from the worst-case size of the dia-
file-searching problem. Another impor- gram. In d >3 dimensions, the pro-
tant application stems from a common hibitively large size of the diagram may
data classification rule that requires as- be reduced to some extent for that
signment of new data points to the same purpose [Dewdney 1977].
class in which their nearest neighbor lies. Generalizations of the post-office prob-
An early solution yielding O(log n) query lem for nonpoint sites or for modified
time but requiring 0( nz) storage was distance functions may be solved by
provided by Dobkin and Lipton [19761. means of an appropriate Voronoi dia-
Shames [19751 pointed out that the gram as well. We particularly mention
Voronoi diagram of the sites divides the Mount [1986] (geodesic distance on a
plane into regions of equal answer with polyhedral surface) and Aurenhammer
respect to the post-office problem: By def- et al. [19911 (probabilistic distances in
inition, the region of a site contains all the plane) as two generalizations leading
points closer to this site than to all other to interesting types of diagrams.

2. 1.2 k-Nearest-Neighbor Queries
An obvious and important generalization

of the post-office problem is to ask for the /
k closest sites for a query point. It was 1, /
I 1, /
/
observed by Shamos and Hoey [1975] that /
1’ /
the order-k Voronoi diagram divides the \
plane into regions of equal answer with \ v’
respect to this problem. Postprocessing <“ ‘1 i>

----
the diagram for point-location yields a --- +“ =..
l’, ‘.
data structure that handles queries in -.
O(log n + k) time and requires O(k2 n) @ I \
I 1
storage. [The diagram may contain \
; \
@(k( n – k)) regions for each of which 1
the k closest sites are stored.] If k is /’ I
I
specified in the input, an 0( n3)-space
data structure yielding the same query Figure 23. All closest pairs and the largest empty
time was obtained by Edelsbrunner et al. circle.
[19861 via embedding the whole family of
order-k Voronoi diagrams for the sites
into an arrangement of planes. A tech-
nique for compacting order-k Voronoi di-
agrams recently developed by Aggarwal
ger of collision, In other applications, for
et al. [19901 drastically reduces the stor-
example, in clustering, one needs to find
age requirement of these approaches [to
the closest site for each of the sites. The
0(n) for fixed k and to 0( n log n) if k is
straightforward approach clearly results
in the input] while still achieving the
in calculating 6( nz) distances. In the
optimal O(log n + k) query time.
planar case, both questions were settled
A related problem is the circular re-
in optimal time 0( n log n) by Shamos
trieval problem that requires a report of
and Hoey [19751 by precomputing the
all sites covered by a given query disk.
Voronoi diagram of the sites. See Figure
In facility location, for example, one
23 depicting all closest pairs by bold seg-
might be interested in all sites being
ments. Since the region of a site and the
influenced by a new facility of given ra-
region of its closest site always share an
dius of attraction. First steps toward a
edge, one only needs to visit each edge
satisfactory solution were undertaken by
and calculate the distance between the
Bentley and Maurer [19791. Their method
two sites whose separator yields that
is based on a certain sequence of order-k
edge. This clearly takes 0(n) time once
Voronoi diagrams of the given sites and
the diagram is available. Recently,
has been refined by Chazelle et al.
Hinrichs et al. [1988a, 1988bl showed
[1986b] and recently by Aggarwal et al.
that it suffices to maintain certain parts
[19901. The latter refinement yields a
of the diagram during a plane sweep
data structure using 0( n log n) space and
rather than to compute the diagram ex-
achieving O(log n + t) query time, where
plicitly in order to solve both problems in
t is the number of sites in the query disk.
optimal time. Curiously, the furthest pair
of sites cannot be derived directly from
2. 1.3 Closest Pairs the furthest site Voronoi diagram but
requires some additional point-location
Finding the closest pair among n sites [Toussaint and Bhattacharya 19811.
certainly belongs to the most fundamen- Finding the furthest pair means de-
tal proximity problems. One obvious ap- termining the diameter of the under-
plication arises in collision detection; the lying set of sites and thus has several
two closest sites are in the greatest dan- applications.

2.2 Placement and Motion Planning

\/
Voronoi diagrams have the property that
each point on a Voronoi edge maximizes
the distance to the closest sites. Al-
though being straightforward, this prop-
erty is essential for the usefulness of
Voronoi diagrams in placement and
motion planning problems.
2.2.1 Largest Empty Figures
The largest empty figure problem is the /“

\
/
following: Given n point sites in the \
plane, where should a figure of pre- \
scribed shape be placed so that its area is Figure 24. Furthest site Voronoi diagram and
maximized but no site is covered? If the smallest enclosing circle.
figure is a circle, the problem can be
reformulated as finding a new site being
as far as possible from n existing ones. a considerable literature [Preparata
Shames and Hoey [1975] mentioned the and Shames 1985]. In particular, the
relevance of this facility location problem problems are solvable in 0(n) time by
to operations research and industrial en- means of linear programming as was
gineering. They also showed how to find demonstrated by Megiddo [1983].
a largest empty circle in 0( n log n) time: The Voronoi diagram approach gener-
The center of such a circle is either a alizes—while still achieving the
vertex of the Voronoi diagram of the sites 0( n log n)-time bound—to other figures
or the intersection of a Voronoi edge and provided these are “circles” with respect
the boundary of the convex hull of the to the underlying distance function. For
sites; see Figure 23 illustrating the oc- example, the largest empty axis-parallel
currence of the latter case. Thus atten- square can be found using the Voronoi
tion can be restricted to 0(n) possible diagram in the L~-metric. Chew and
placements of the circle, Drysdale [1985] find largest empty
The smallest enclosing circle of the homothetic copies of generally shaped
sites is obtainable in 0( n log n) time in a convex figures via diagrams that are
similar manner. Shames [1978] claimed induced by a rather general class of
that this circle always will be centered at distance functions. The determination of
a vertex of the furthest site Voronoi dia- the largest empty axis-parallel rectangle
gram. This was proved true by Toussaint turns out to be somewhat more compli-
and Bhattacharya [19811. From Figure cated; the aspect ratio of the rectangle,
24 it can be seen that the edges of this and thus the distance function to be used,
diagram form a tree structure. This re- is not known in advance. This problem
stricts the maximum number of vertices, has an obvious application. If we are
and thus of placements of the circle, to given a piece of fabric or sheet metal and
n – 2.7 On both problems, there exists the sites are flaws, what is the largest
area rectangular piece that can be sal-
vaged? Chazelle et al. [1986a] gave an
0( n log3 n)-time solution. They use an
7 It is easily seen that the regions of the furthest interesting type of Voronoi diagram that
site diagram are either unbounded or empty, The can be intepreted as the power diagram
corresponding furthest site Delaunay triangulation
of a certain set of circles [Edelsbrunner
thus is an outer planar graph (with at most n sites
as vertices), which is well known to contain at most and Seidel 1986; Aurenhammer and Imai
2 n – 3 edges and n – 2 triangles. 1988]. The running time above was
ACM Computing Surveys, Vol. 23, No, 3, September 1991

improved to 0( n log n) by Aggarwal and ,- .-, _---–

Suri [1987] using different methods. The / ~—. > -——- -——
properties of the Voronoi diagram for line /’
/,’/
/.’
//
segment sites have been exploited by ;-
/“
L..
Chew and Kedem[1989] to find the largest /’
.— ------ ---
-_
/’ ---
similar copy of a convex polygon such /
/’ /
that no segment is overlapped. As a re- .
lated question, Aonuma et al. [1990] place . /’

a given convex polygon inside and as
far as possible from the boundary of an Figure 25. Voronol diagram for line segments.
arbitrary polygon. Their algorithm relies

on Voronoi diagrams dynamized with
respect to certain site movements. the clearance falls short of the disk
radius. All of that can be done in
2.2.2 Translational Motions additional 0(n) time.
Compared to others, this simple ap-
Voronoi diagrams are an aid in the plan- proach has the additional advantage that
ning of collision-free motions of a figure the disk radius need not be known dur-
in the presence of obstacle sites. Prob- ing the preprocessing phase. Extending
lems of this kind arise in robotics. The the approach, Leven and Sharir [19871
Voronoi diagram approach plays an im- plan translational motions of a convex
portant role in this area and is called the planar robot in 0( n log n) time. Note
retraction method. Intuitively speaking, that generally shaped robots may be cir-
when moving on the edges of the Voronoi cumscribed by a convex figure, and paths
diagram the robot always keeps as far feasible for the figure will do for the
as possible from the neighboring ob- robot as well.
stacles. This was first perceived by
Rowat [19791; several authors pursued 2.2.3 Rotational Motion
this idea further.
For many applications it is feasible to Allowing the robot to rotate adds a third
approximate the scenario of obstacles by degree of freedom to the problem. Much
polygonal sets. O’Dunlaing and Yap attention has been paid to moving—
[19851 plan the motion of a disk between amidst polygonal obstacles— a line
two given points in this environment. In segment that is allowed to rotate. The
a preprocessing step, the Voronoi dia- position of a segment (of given length) in
gram of the obstacles is computed by tak- the plane is determined by the triple p =
ing their n boundary segments as its (x, y, p) indicating the coordinates of one
defining sites. The graph formed by the endpoint and the orientation. The set of
0(n) straight or parabolic edges of the all triples p such that the corresponding
diagram can be computed in 0( n log n) segment L(p) does not collide with any
time [Kirkpatrick 1979; Fortune 19871. obstacle is a subset of the underlying
Figure 25 depicts the graph (dashed) for three-dimensional configuration space.
four line segment obstacles (bold). Ob- Within this subset, we may define a
serve that each point on an edge is Voronoi diagram as the set of all p such
equidistant from its two closest obsta- that -L(p) is equidistant from its two
cles. Therefore, for each edge, it is easy closest obstacles; closeness is with re-
to compute the minimum clearance when spect to the minimum distance between
moving on it. To produce a collision-free segment and obstacle.
motion (if it exists), one first finds the O’Dunlaing et al. [1986, 1987] pro-
edges e and e’ closest to the center of the vided a thorough analysis of this dia-
disk in initial and final position, respec- gram (which seems to belong to one of
tively, and then searches the graph start - the most complex types). In particular,
ing with e until either e’ is reached or they showed that the diagram can be

constructed, and a motion be planned in
----
--/.
9
0( n2 log n log* n) time, thus being a sig- -- T-.
nificant improvement over earlier tech-
niques. Sifrony and Sharir 11986] speeded ‘\ ‘\
‘. .
,/- ----
\ -. -.%
up the method—by introducing a related /’
-..
graph-to run in O(t log n) time, where . .
t = 0( n2) is a parameter dependent on
the segment length. Canny and Donald
[19881 simplified the method by relaxing
the definition of the diagram. Figure 26. Geodesic separator.
There are several other versions of the
motion planning problem to which and a collection of disjoint and con-
Voronoi diagrams apply [Rohnert 19881. vex polygons with a total of m edges in
An axiomatic characterization of the 0( n + m)log( n + m) time? The main
boundary properties of a generally shaped problem in this respect is that com-
three-dimensional scenaro in order to de- puting geodesic distances is not a
fine a Voronoi diagram suited to motion constant-time operation, Figure 26 shows
planning was proposed by Stifter [19891. the separator (dashed) of two sites under
For a more detailed introduction and re- the geodesic distance among two convex
view of the use of Voronoi diagrams in polygons (bold). It is composed of straight
motion planning, the interested reader line and hyperbola segments meeting at
may consult Schwartz and Yap [1986] or intersections with visibility lines (solid).
Alt and Yap [19901. Path-planning problems in three-
dimensional polyhedral spaces have been
2.2.4 Path Planning studied, by among other authors, Mount
[19861, Mitchell et al. [1987], and
A special case of planning collision-free Baltsan and Sharir [1988]. For example,
motions, where the robot is taken to be a finding shortest paths lying entirely on a
single moving point, is path planning. (generally nonconvex) polyhedral surface
Finding euclidean shortest paths in the is of interest in areas like autonomous
presence of polygonal (or polyhedral) ob- vehicle navigation, where hilly terrains
stacles has received particular interest. are being modeled. Efficient solutions are
For example, consider the problem of obtained by using Voronoi diagrams de-
determining that location, for each of m fined in such spaces. Moreover, since
fixed locations within a polygonal factory finding shortest paths is closely related
floor, reachable on the shortest path. This to determining visibility between obsta-
problem can be solved in 0( n + cles, Voronoi diagrams involving visibil-
m)log2 ( n + m) time by first constructing ity constraints have been exploited in
the Voronoi diagram of n sites under the this context. An example is the peeper’s
geodesic metric interior to a simple poly- Voronoi diagram used by Baltsan and
gon with m edges [Aronov 1989]. The Sharir [1988] and by Aronov [1989]. Each
site yielding the shortest path to a par- point in the plane is assigned to the clos-
ticular site defines a neighboring region est visible site, and visibility is con-
in the diagram. This improves an earlier strained to a segment on a line avoiding
result by Asano and Asano [19861. Re- the convex hull of the sites (see Figure
lated questions can be solved by means of 36). This structure can attain a size of
the furthest site geodesic Voronoi dia- @ n2) and is constructible in 0( TL2) time
gram inside a polygon; Aronov et al. [Aurenhammer and St6ckl 1988].
[1988] showed how to construct this
structure efficiently. 2.3 Triangulating Sites
The following important question is
still unsolved. Is it possible to compute Intuitively speaking, triangulating n
the geodesic Voronoi diagram for n sites sites in the plane means partitioning the

plane into triangles whose vertices are

all and only the sites. Triangulations are
important because they (as being planar
graphs) on the one hand have only linear
size and, on the other hand, establish
connectivity information among the sites
that suffices for many applications.
2.3.1 Equiangular Triangulations
An early and major application is the

interpolation of functions of two vari-
ables, where function values initially are
known only at irregularly placed sites.
Practical evidence stems from finite
element methods or from processing
graphical data, for example, in terrain
modeling. Given a triangulation of Figure 27. Constrained Delaunay triangulation.
the sites, the function value at an arbi-
trary point can be computed by interpo-
lation within the triangle containing that
minimized if the underlying triangula-
point. Lawson [1972] and McLain [1976]
have reported that a triangulation is well tion of the sites is Delaunay. This re-
sult is somewhat surprising since the
suited to interpolation if its triangles are
nearly equiangular. The Delaunay trian- Delaunay triangulation–although itself
gulation is the unique triangulation that clearly being independent of the height
is optimum in this sense [Sibson 1977]. at each site —optimizes a quantity that
Sibson [19801 further suggests the use of depends on these heights.
local coordinates obtained from sites be-
ing adjacent in this triangulation in or- 2.3.2 Constrained Triangulations
der to compute smooth surfaces. There In a constrained Delaunay triangulation,
might, however, occur more-sided poly- prescribed edges are forced in as part of
gons (rather than triangles) if various the triangulation. See Section 3.3 for a
sites are cocircular. Mount and Saalfeld formal definition. Figure 27 depicts such
[19881 showed how to triangulate such a triangulation with two prescribed edges
polygons in order to retain equiangular - (bold), This structure can be constructed
ity. This process does not affect the 0( n in 0( n log n) time [Seidel 1988] and has
log n) time needed for constructing the two major applications. First, it provides
Delaunay triangulation. a more realistic approach to modeling
Given a set of sites and a particular terrain surfaces since salient elements
function value (height) at each site, any like mountain ridges or valleys may be
triangulation of the sites defines a trian- prescribed. Lee and Lin [1986] showed
gular surface in space. The “roughness” that the equiangularity criterion is still
of such a surface T may be measured by fulfilled by constrained Delaunay trian-
the Sobolev seminorm gulations, Since triangles with small an-
gles produce a poor computer graphics
display, the best possible visualization is
achieved. Triangulations inside a simple
where I A I denotes the area of the spa- polygon are an important special case.
tial triangle A, and a~ and 6A denote the The boundary edges of the polygon will
slopes of the plane containing A. By ex- not be edges of the (unconstrained)
ploiting equiangularity, Rippa [1990] Delaunay triangulation of the polygon
showed that the Sobolev seminorm is vertices, in general, and thus have

to be prescribed.8 Chew [1989b] iteratively adds shortest diagonals by

used constrained Delaunay triangula- considering the current (and initially
tions to generate triangular meshes empty) set of edges as prescribed. Lingas
inside polygons where all angles are [19891 reports that a shortest diagonal
between 30° and 120°. can be found in 0(n) time given the
Second, constrained Delaunay triangu- constrained Delaunay triangulation. This
lations are an aid in path planning. In gives an 0( n2 log n)-time and 0( n)-space
order to plan a path from A to B that algorithm for constructing such greedy
avoids a set of n line segment obstacles triangulations. Previous methods re-
(and thus arbitrarily shaped polygonal quired 0( n2) space. Whether a minimum
regions, for instance) one might con- length triangulation for n sites in the
struct this triangulation-by taking A plane can be computed in polynomial
and B and the segment endpoints as sites time is an important open question.
and the segments as prescribed
edges— and then search the graph formed
by triangulation edges. This was ob- 2.3.3 3D Triangulations
served by Chew [19861, who also proved
The problem of “triangulating” point
the following interesting result: Between
sites in three-space arises in the decom-
any two sites there is a path in this position of three-dimensional objects, for
triangulation whose length is at most
example, of computer-generated solid
~ times the geodesic distance between models or of objects specified by planar
them, provided the L1-metric (instead of cross sections. Field [19861, Boissonnat
the euclidean) is taken to define the tri- [1988], and others report on advantages
angulation. This yields an 0( n log n)- of tetrahedral decompositions and pro-
time algorithm for approximating the pose the use of the Delaunay triangula-
optimal path, which is a significant tion. Although there is some evidence for
improvement over exact methods. the quality of thus produced tetrahedral,
Dobkin et al. [1990] proved a factor of some important properties of the two-
about 5 for (nonconstrained) euclidean dimensional Delaunay triangulation do
Delaunay triangulations and the eu- not carry over.
clidean distance. Their result has been First, the number of tetrahedral maybe
improved to a factor of about ~ by Keil E)( n2), hence leading to a storage re-
and Gutwin [19891. Results of this kind quirement that is prohibitive for many
are rather surprising since they show the applications. This fact suggests the use
existence of a sparse graph being almost of postprocessing. ~hazelle et al. [1990]
as “good” as the complete graph on n showed the following somewhat counter-
sites, independently of n. intuitive result: The number of tetrahe -
Let us mention another application. dra may always be lowered to about n G
Given n sites and some prescribed (non-
by adding about & new sites. This pro-
crossing) edges between them, a diagonal
cess will also eliminate elongated (and
is a new edge that does not cross any
thus undesirable) tetrahedral, In fact,
prescribed one. A popular heuristic for
adding 0(n) (well-chosen) sites will al-
triangulations with minimum total edge
ways reduce the size of the triangulation
length is the greedy method. This method
to 0(n), with the additional effect that
only constantly many edges emanate
from each site [Bern et al. 19901.
6 It is an interesting question how many new sites Second, the question of finding a
have to be placed on the polygon boundary in order
three-dimensional analog to the equian -
to force it in as a subgraph of the Delaunay trian-
gulation. Schwarzkopf (personal communication, gularity property is still unsettled. Rajan
1990) observed that !l(nz) sites may be necessary [19911, however, recently succeeded in
for an n-vertex polygon. A known upper bound is proving a related property. The three-
2 o(m log ~), The O(~) bound ~laimed in Boissonnat
dimensional Delaunay triangulation
[19881 involves a constant depending on the angles
of the polygon.
minimizes the largest containment

radius of the tetrahedral; that is, the ra - straight-line connection of the sites with
dius of the smallest sphere containing a minimum total edge length. See Figure 8
tetrahedron. This sphere coincides with for an illustration. Not surprisingly, this
the circumsphere if and only if its center structure plays an important role in
lies within the tetrahedron. Note that transportation problems, pattern recog-
the containment radius measures the te- nition, and cluster analysis. Construc-
trahedron size more appropriately than tion methods working on general
the circumradius since small but elon- weighted graphs have been known for a
gated tetrahedral may have arbitrarily long time; see Kruskal [19561, Prim
large circumradii. The Delaunay trian- [19571, and, more recently, yao [19751
gulation is the most compact triangu- whose algorithm works in 0( m log log n)
lation in this sense. This property is time for a graph with m edges. All these
quite remarkable as—unlike the two- algorithms use a basic fact: For any par-
dimensional situation—the number of te- tition of the sites into two subsets, the
trahedral is not independent of the way of shortest edge (edge of minimum weight)
triangulating the given sites but may between the subsets will be present in
vary from linear to quadratic. the tree. Applying Yao’s algorithm di-
There is another property of Delaunay rectly to the complete euclidean distance
triangulations that turns out to be useful graph for the sites results in an Q( nz)
in three-dimensions. For a set of trian- running time, however.
gles in three-space, an in-front/behind Shames and Hoey [19751 recognized
relation may be defined with respect to a that the edges of a minimum spanning
fixed viewpoint. Generalizing a result by tree must be Delaunay edges. This inter-
De Floriani et al. [19881, Edelsbrunner esting and important property holds
[19891 proved that this relation is acyclic since, by the fact mentioned above, each
for the triangles in a three-dimensional tree edge is the diameter of a disk empty
Delaunay triangulation, no matter where of sites and, by definition, each triangu-
the viewpoint is chosen. This is relevant lation edge is a chord of some disk empty
to a popular algorithm in computer of sites. From the computational view-
graphics that eliminates hidden surfaces point, Yao’s algorithm just needs to be
by first ordering the three-dimensional applied to the Delaunay triangulation— a
objects with respect to the in-front/ graph with 0(n) edges. In fact, 0(n) time
behind relation and then drawing them suffices to derive the minimum spanning
from back to front, thus over-painting tree from the Delaunay triangulation
invisible parts. For Delaunay triangles, [Preparata and Shames 19851. This gives
such an ordering always will exist, which an 0( n log n)-time algorithm, which is
is quite important in view of their optimal by reduction to sorting n real
frequent use in practice. In particular, numbers.9
so-called a-shapes are made of such The relationship above extends to more
triangles. a-shapes have been used by general metrics. For example, Hwang
Edelsbrunner et al. [19831 to model the [1979] extends it to the L1-metric (com-
shape of point sets. pare Figure 3). One may be tempted to
conjecture that the (euclidean) maximum
2.4 Connectivity Graphs for Sites spanning tree can be obtained from the
furthest site Delaunay triangulation of
The Delaunay triangulation contains, as the sites (the dual of the furthest site
subgraphs, various structures with far-
ranging applications. We will briefly
discuss some of them in the sequel.
9 The minimum spanning tree should not be con-
fused with the minimum Steiner tree of a set of
2.4.1 Spanning Trees
sites, a concept where the addition of new points is
allowed in order to minimize the total edge length.
Given n point sites in the plane, a The problem of constructing a minimum Steiner
(euclidean) minimum spanning tree is a tree is NP-complete [Garey et al. 1976].

Voronoi diagram). Shames [1978] showed

that this conjecture is erroneous.
Eppstein [1990], however, pointed out
that two sites are connected in one of
the k smallest spanning trees only if
both belong to the same subset of sites
defining a region of the order-(k + 1)
Voronoi diagram. This leads to an
improved algorithm for finding the k
smallest spanning trees.
The efficient construction of the mini-
mum spanning tree in three space has
turned out to be a diffkult problem. Pre -
computation of the Delaunay triangula- Figure 28. Delaunay triangulation and salesman
tion does not help in general since the tour [Dillencourt 1987].
complete graph may be the output in the
worst case. Efficient approximation and
expected-time algorithms were given,
among other authors, by Vaidya [1988a] sites. Constructing traveling salesman
and Clarkson [1989]. The first algorithm tours is a prominent problem of combina-
running in subquadratic worst-case time torial optimization. Papadimitriou [1977]
is due to Yao [1982]; it can be speeded up showed its NP-completeness. The prob-
to run in 0( n log n)15 time by using lem becomes tractable for tours being
Voronoi diagrams of carefully chosen close to optimal. It follows from the above
grOUpS of sites. By exploiting—in addi- discussion that the Delaunay triangula-
tion—an interesting relation to a certain tion will not contain a traveling sales-
closest point problem, Agarwal et al. man tour, in general. 10 See also Fig-
[19901 were able to bring down the com- ure 28 where the triangulation and the
plexity to 0(n413 log4/3 n). It still re- tour are drawn solid and dashed, re-
mains open whether an algorithm spectively. Rosenkrantz et al. [19771,
matching the best currently known lower however, observed that traversing the
bound of !2( n log n) can be developed. minimum spanning tree twice will pro-
duce a tour that is within a factor of
2.4.2 Spanning Cycles
2 of the optimum: Since removing an
edge from the traveling salesman tour
A graph connecting n sites is called leaves a spanning tree of the sites, this
Hamiltonian if it contains some cycle tour must be longer than the minimum
passing through all the sites. The ques- spanning tree. Note that this approxima-
tion of whether the Delaunay triang-ula- tion algorithm takes only 0( n log n)
tion in the plane is Hamiltonian arose in time. By partitioning the minimum
pattern recognition problems where spanning tree into paths, then construct-
a reasonable simple curve through ing a minimum-length matching of their
the given sites is desired. Although endpoints, Christofides [1976] improved
counter-examples were given by this to a factor of 1.5, with the expense of
Kantabutra [19831 and Dillencourt an 0( n2 A log4 n) construction time due
[1987a, 1987b], Delaunay triangulations
have been used with success for this
problem. This gives evidence that this
graph is Hamiltonian with high proba-
bility. Dillencourt [1987cI supports this 10 This is even true for special traveling salesman
tours called necklace tours that are cycles realiz-
thesis by proving a result concerning the
able as the intersection graph of disks around the
connectivity of Delaunay triangulations. sites. Necklace tours can be found in polynomial
A (euclidean) traveling salesman tour time in case of their existence [Edelsbrunner et al.
is a minimum-length cycle spanned by n 1987].

fortunately, no algorithm with signifi-

cantly subquadratic running time is
known.
A similarly defined construct is the
Gabriel graph. It contains an edge be-
tween two sites if the disk having that
edge as its diameter is empty of sites.
This concept proved useful in processing
geographical data [Gabriel and Sokal
1969; Matula and Sokal 19801. Howe
[1978] observed that the Gabriel graph
Figure 29. Relative neighborhood graph and consists of just those Delaunay edges that
Gabriel graph. intersect their dual Voronoi edges. Based
on the same observation, Urquhart [19801
gave an 0( n log n) time algorithm for its
to the best known matching algorithm by construction. It is easily seen that the
Vaidya [1988bl.11 Gabriel graph is a supergraph of the rel-
ative neighborhood graph. The edges of
2.4.3 Relative Neighborhood and Gabriel Graph the former are shown solid or dashed in
Figure 29.
It is worth mentioning two graphs that
are “in between” the minimum spanning
2.5 Clustering Point Sites
tree and the Delaunay triangulation in
the sense of making explicit more prox- Clustering a set of n point sites in the
imity information than does the former plane means determining partitions of
and less than does the latter. Among the set that optimize some predefine
them is the relative neighborhood grap~ clustering measure. This measure usu-
that connects two sites provided no other ally is a function of the interpoint
site is closer to both of them than their distances of the set to be clustered.
interpoint distance. In Figure 29, the
edges of the relative neighborhood graph
are shown solid. Toussaint [19801 re - 2.5.1 Hierarchical Methods
ported the usefulness of this structure in A single linkage clustering [Hartigan

pattern recognition, and Supowit [19831 19751 hierarchically clusters the sites as
succeeded in constructing it in 0( n log n) follows. Initially, the sites are considered
time, given the Delaunay triangulation. to be clusters themselves. As long as
Yao (personal communication, 1988) an- there is more than one cluster, the two
nounced that 0(n) time suffices for de- closest clusters are merged. The distance
riving the relative neighborhood graph of two clusters is defined as the mini-
from the latter. Interestingly, this graph mum distance between any two sites, one
retains its linear size in higher dimen- from each cluster. This process can be
sions (provided a general placement of carried out in a total of 0( n log n) time
the sites). This stresses the importance of by maintaining, at each stage, the fol-
its rapid construction, especially in three lowing Voronoi diagram: Each point of
dimensions, since it would imply fast the plane belongs to the region of the
minimum spanning tree algorithms. Un- closest cluster, and closeness is with re-
spect to the closest site within a cluster.
The regions of the two closest clusters
11A minimum-length matching of 2 n sites is a will always have a common edge in the
graph with n edges joining pairs of sites, such that diagram. That edge corresponds to an
the pairs are distinct and the total edge length is
edge of the minimum spanning tree of
minimized. Akl [19831 showed that a minimum-
length matching is not necessarily a subgraph of the entire set of sites [Shames and Hoey
the Delaunay triangulations of the sites. 1975].

lution was proposed by Asano et al.

[19881. Construct the minimum spanning
tree of the sites and remove its t – 1
/
/ longest edges. The resulting t subtrees
/ already give the desired clustering. Us-
/
\\ ing similar methods, Asano et al. [19881
., obtain an equally efficient algorithm for
\
\/’/\ partitioning the sites into two clusters so
\T\\
‘\
that their maximum diameter is mini-
\ mized. This problem becomes consider-

ably harder if partitions into more than
two clusters are sought. 12 Partitions into
separable clusterings have been investi-
I ‘. gated by Dehne and Noltemeier [19851
and Heusinger and Noltemeier [19891,
Figure 30. Complete linkage Voronoi diagram. who exploited the fact that clusters of k
sites being separable from the remaining
sites by a straight line define unbounded
The model of complete linkage cluster- regions in the associated order-k Voronoi
ing is similar, except that the distance of diagram.
two clusters is defined as the maximum
2.5.3 Optimum Cluster Selection
distance between any pair of sites, one
from each cluster. The corresponding The k-variance problem asks for select-
Voronoi diagram, where closeness of a ing, from a given set S of n sites, a
point in the plane is with respect to the cluster of k sites with minimum vari-
furthest site in a cluster, deserves special ance, the sum of squares of all intersite
attention. Its size is fI( rzz) for general distances in the cluster. Aggarwal et al.
clusters and 0(n) provided the convex [1989b], who cite applications in pattern
hulls of the clusters are pairwise disjoint recognition, observed the following inter-
[Edelsbrunner et al. 1989]. Figure 30 esting fact. If C is k-cluster of minimal
shows the diagram induced by four two- variance, then the region of C in the
site clusters, indicated by bold segments. order-k Voronoi diagram of S is
For each cluster, the portion of the bisec- nonempty, Thus it suffices to examine
tor of its two sites that lies within its the variances of sets of k sites associated
region is drawn dashed, The diagram can with Voronoi regions. This can be done
be constructed in 0( n log n) random- in time proportional to the number,
ized time for convex-hull disjoint and O(k(n – k)), of regions by exploiting the
constant-sized clusters and— in this fact that sets of neighbored regions differ
setting— applies to the retrieval of in exactly two sites. Thus, the most
expected-nearest sites [Aurenhammer time-consuming step is the precomputa-
et al. 1991]. Efficient maintenance of the tion of a higher order Voronoi diagram.
diagram during the clustering process A similar approach leads to improve-
would lead to an improved clustering al- ments if the cluster measure to be mini-
gorithm but remains open. mized is not variance but diameter. If C
is a k-cluster with minimum diameter,
2.5.2 Partitional Strategies then C is contained in some set of 3k – 3
sites whose region in the order-(3k – 3)
As another example, consider the prob- Voronoi diagram of S is nonempty;
lem of partitioning a set of n sites into a
fixed number of t clusters such that the
minimum of the (single linkage) dis- 12 Asano et al. [1988] claimed its NP-completeness.
tances between clusters is maximized. Rote (personal communication, 1990), announced a
The following elegant 0( n log rz)-time so- polynomial algorithm in the case of three clusters.

see Aggarwal et al. [1989bl where

an 0(k2 n ~k log n)-time selection
algorithm is obtained in this way.
Several other clustering algorithms
that rely on Voronoi diagrams have been
described in the literature; some of them
are mentioned in the Introduction.
3. SELECTED TOPICS
A thorough understanding of the geomet -

ric, combinatorial, and topological prop-
erties of Voronoi diagrams is crucial for
the design of efficient construction algo-
rithms. This third part of the survey
presents these properties in a unified
manner and discusses their algorithmic
implications. We start with a description
of the various relationships of Voronoi
diagrams to objects in one more dimen- Figure 31. Power line of two circles.
sion. We continue with some topological
properties that are particularly useful for
the divide-and-conquer construction of a
large class of Voronoi diagrams in the individual real number w(p) called its
plane. Finally, we investigate a continu- weight. ‘I’he power function of a point
ous deformation of the planar Voronoi x e Rd with respect to p is specified by
diagram being well suited to construc-
tion by the plane-sweep technique
and extend it to generalized Voronoi pow(x, p) = (.x- p) T(x -p) - w(p).
diagrams.
For W(p) >0, one may think of the
3.1 Geometry of Voronoi Diagrams: Their
weighted site p as a sphere in R d with
Relationship to Higher Dimensional
Objects
radius - around p; for a point x
outside this sphere, pow( x, p) > 0, and
Voronoi diagrams are intimately related v“- expresses the distance of x
to several central structures in discrete to the touching point of a line tangent to
geometry. This section exhibits these re- the sphere and through x. In the plane,
lationships in a comprehensive manner the locus of equal power with respect to
and discusses some of their geometric two weighted sites p and q is commonly
and algorithmic consequences. Although called the power line of the correspond-
our main interest is in the classical type ing circles. If the circles are nondisjoint,
of Voronoi diagram, it is advantageous to their power line passes through their
base the discussion on the more general points of intersection; see Figure 31. In
concept of power diagram since the geo- general dimensions, the locus of equal
metric correspondences to be described power is a hyperplane in R‘, called the
extend to that type in a natural way. We chordale of p and q or chor( p, q), for
shall refer to d dimensions in order short. chor( p, q) is orthogonal to the
to point out the general validity of the straight line connecting p and q but
results. does not necessarily separate p from q,
For W(p) = w(q), chor(p, q) degenerates
3. 1.1 Convex Polyhedral Surfaces
to the symmetry hyperplane of p and q.
Consider a set S of n point sites in R d Let h( p, q) denote the closed halfspace
such that each site p G S is assigned an bounded by chor( p, q) and containing the

Figure 32. Spatial interpretation of chordale [Aurenhammer 1987a].
points of less power with respect to p. thehyperplane

The power cell of p is given by
7r(p): Xd+l = 2xTp –pTp + w(p)
cell(p) = ,E~~fP}h(p, q).
in R‘+ 1 [Aurenhammer 1987a]. This fact
is best explained by reference to Fimre
In analogy to classical Voronoi regions, 32. Let p and q -be two sites in- R2
the power diagram of S, PD(S) is the whose weights are indicated by circles.
convex polyhedral complex defined in R d Projecting the circles onto the paraboloid
by these cells; see Figure 4 for the case of revolution xs = x; + X2 2 defines two
d = 2. PD(S) coincides with the classical planes T(p) and T(g) that intersect the
Voronoi diagram of S in the case of paraboloid in these projections. It is an
equally weighted sites. Note that power easy analytical exercise to prove that the
cells may be empty if general weights line T(p) fl x(q) projects orthogonally to
are used. the chordale chor( p, q) in the xl xz-plane.
For technical reasons let us identify This correspondence holds for arbitrary
R d with the linear subspace of Rd+ 1 dimensions. As an immediate conse-
orthogonal to the (d + l)s~ coordinate quence, PD(S) is the orthogonal projec-
axis xd+ ~. The key observation for the tion of the boundary of the polyhedron
existence of objects in R‘+ 1 that are re. that comes from intersecting the half-
lated to power diagrams is that the power spaces of Rdw~ above m(p), for all p c S.
function pow( x, p) can be expressed by In fact, m is a bijective mapping between
ACM Computing Surveye, Vol. 23, No. 3, September 1991

weighted sites in R d and hyperplanes in Aurenhammer [1987a]. To this end, the

R‘+ 1. Thus, for any intersection of upper concept of a power cell is generalized to
half-spaces there exists a corresponding more than one site, Let T be a subset of
power diagram in one dimension lower k sites in S. The power cell of T is
and vice versa. defined as
This general result has far-reaching
implications. Most important, it shows cell(T) = o h(p, q).
that power diagrams are in a geometric PGT, q=S– T
and combinatorial sense equivalent to
unbounded convex polyhedra or, more
The complex of all (nonempty) power cells
precisely, to their boundaries, that is, to
convex polyhedral surfaces. Convex poly- arising from the ~ subsets of S with
()
hedra are well-understood objects fixed cardinality k. is known as the
in discrete geometry, hence so are power order-k power diagram of S, k-PD(S).
diagrams. Exact upper and lower bounds Clearly 1-PD(S) = PD(S) holds. Note
on the numbers of their faces of various that k-PD( S) is just the order-k Voronoi
dimensions are known. In particular, diagram of S provided all sites in S are
PD(S) in Rd realizes at most n cells, L equally weighted. ( n – 1)-PD(S) is also
j-dimensional faces (1 = j < d – 1),and called the furthest site power diagram
f. – 1 vertices, for of s.
Now let Z denote the intersection of
n–d+i–2 the half-spaces below the hyperplanes
~=~o(~)( i )
T(p) for all p e T and above the hyper -
planes x-(q) for all q e S – T. Due to
the foregoing reasoning, Z projects or-
‘io(d-~+l)(n-d: i-2) thogonally onto cell(T) in R‘. This
observation draws the connection
between order-k power diagrams and so-
called hyperplane arrangements. The lat -
This follows from the so-called upper ter are cell complexes that arise from
bound theorem [Brondsted 19831. The dissecting the space by a finite number of
numbers ~ are O(n Fdlzl ) for O < j s hyperplanes. Each cell in an arrange-
d–1. ment is the intersection of some half-
As for the facets of a convex polyhe- spaces below or above its defining n hy -
dron, the power cells are convex but pos- perplanes. If we let a k-level consist of all
sibly unbounded polyhedra. Not every cells lying below k and above n – k hy -
half-space within a given set needs to perplanes, we can conclude that the k-
contribute to a facet, so power cells may level induced by {T(p) I p = S} projects to
be empty or degenerate. d + 1 or more the cells of k – PD(S) for each k be-
hyperplanes in R d+ 1 intersectin a com- tween 1 and n – 1.This shows that each
mon point (unless parallelism occurs), so arrangement in R‘~ 1 corresponds to a
at least d + 1 power cells meet at each complete family of higher order power
vertex of PD( S). In particular, there are diagrams in R ~ and vice versa. The cor-
no vertices if the cardinality n of S does respondence between order-1 power dia-
not exceed d. These and many other grams and convex polyhedral surfaces
properties of PD(S) can be read off con- discussed earlier is a special case of this
veniently from its (d + 1)-dimensional situation.
embedding, Figure 33 depicts a cell (left) being the
intersection of five upper and three lower
3. 1.2 Hyperplane Arrangements
half-spaces, that is, a three-level cell. Its
orthogonal projection onto the xl xz-plane
In order to exploit the situation fully, is an order-3 power cell among eight sites
we turn our attention to diagrams of (right) whose boundary is indicated by
higher order [Edelsbrunner 1987]; bold lines. Note that the power cell is

Voronoi Diagrams * 383
e .“
●
.. /
“.
/-
+p~o”
O“’’’’&’’’’: I
i
i
;
i
i
i
●
Figure 33. Projecting an arrangement cell [Aurenhammer 1987a].
carved into subcells by the furthest site We state another interesting conse-
power diagram of the three emphasized quence. Let w’(p) denote the image of
sites and by the power diagram of the reflection of the hyperplane T(p) through
remaining five sites. Rd. The cells in the k-level of { w’( p) I p e
The correspondence to arrangements S} correspond bijectively to the cells in
straightforwardly yields an upper bound the (n – k)-level of {r(p) I p c S} by re-
on the overall size of all higher order versing “upper” and “lower” for their
power diagrams of a fixed set of n sites: supporting half-spaces. Clearly, any two
the maximal number of faces of an ar- corresponding cells yield the same or-
rangement of n hyperplanes in Rd+ 1. thogonal projection onto Rd. That is to
This number is precisely say, they yield the same power cell that
is simultaneously of order-k and order-
(n – k) for fixed k. This shows that any
,=2:+1( d–~+1“ )(:) order-k power diagram for n sites can be
interpreted as some order-( n – k) power
for faces of dimension j, which is in diagram. In particular, the furthest site
O(n~+l) for 0 = j < d + 1 [Alexanderson power diagram of S is the order-1 power
and Wetzel 1978]. Edelsbrunner et al. diagram of some set of sites (distinct from
[1986a] developed an algorithm for con- S, in general). We point out that this is
structing such arrangements in 0( nd+ 1) not true for the subclass of classical
time. Consequently, the complete family Voronoi diagrams, as can be seen from
of higher order power (or Voronoi) dia- the maximal size of their closest site and
grams of a given set of sites can be furthest site counterparts that differ even
computed in optimal time and space. in R2; see the Introduction and Section
Edelsbrunner [1986] showed how to con- 2.2.1.
struct the k-level of an arrangement Finally, let us mention a result on
in R 3, and thus a single order-k power cells in R2. Let T c S and I T I =
power (or Voronoi) diagram in the plane, k. By definition, cell(T) is the intersec-
at a cost of 0(6 log n) per edge. tion of k( n – k) half planes and thus
Recently, Mulmuley [19911 proposed a might have that many edges. On the
randomized algorithm for constructing other hand, cell(T) is the projection of a
levels of order 1 to k in d >4 dimen- polyhedron in R3 that has at most n
sions in worst-case optimal time. facets and thus only 0(n) edges by

Euler’s relation. It follows that cell(Z’)

\ /
has O(n) edges and can be constructed
in 0( n log n) time by intersecting n
half-spaces of R3.
3. 1.3 Convex Hulls
The convex hull of a finite point set is

the intersection of all half-spaces con-
taining this set. Next, we study the dual-
ity of (order-1) power diagrams and
convex hulls [Aurenhammer 1987al.
Duality is usually defined for con-
vex polyhedra. Let Z and Z’ be
two convex polyhedra in R‘+ 1. Z and Z’
are said to be dual if there is a bijective Figure 34. Polarity between points and lines.
mapping ~ between the j-dimensional
faces of Z and the (d – j)-dimensional
faces of Z’ such that f G g, for any two
q~s. This, however, means that
faces f and g of Z if and only if ~(g) ~
‘(PI), . . . , ~(P~) span a (d – j)-
4(f). Since power diagrams in Rd are dimensional face of the 10wer part of the
projections of convex polyhedral surfaces
convex hull of the point-set {A(p) I p e S}.
in Rd+ 1, the notion of duality carries
This lower part consists of all boundary
over in a natural manner. faces of the hull that are visible from the
We now consider, for each site p E S
point on the xd+ ~-axis at – co. In other
with weight W(p), a point in Rd+ 1:
words, we have shown the duality of Z,
and thus of PD(S), to that lower convex
(
P
A(p) = hull part.
PTP – W(P) ) By the same reasoning, the duality of
the furthest site power diagram of S and
From the formula expressing the hyper- the upper part of the convex hull of
plane T(p) itcan be seen that A(p) and {A(p) I p e S} can be established. This
T(p) are related via polarity with re- reveals another interesting link between
spect to the paraboloid of revolution. A(p) closest site and furthest site power dia-
is the pole of m(p) and x(p) is the polar grams and, particularly, classical Voronoi
hyperplane of A(p). If T(p) happens to diagrams. Most important, these dia-
intersect the paraboloid, then polarity grams become constructible in general
has the geometric interpretation dis- dimensions d via convex-hull algo-
played in Figure 34. The hyperplanes rithms. Running times of 0( n log n) for
touching the paraboloid at its points of d = 2 and of 0(nrdi21 ) for odd d are
intersection with 7(p) all concur at A(p). achieved, both of which are worst-case
Recall that the intersection, Z, of the optimal; see Preparata and Hong [19771
half-spaces above ~( p), for all p e S, pro- and Seidel [1981], respectively. For even
jects to PD(S). Take an arbitrary d >4, these diagrams can be computed
j-dimensional face of Z that lies in the at logarithmic time per face [Seidel 1986].
intersection of, say, T( pi), . . , 7r(pn), By projecting down the edges of the
m 2 d —~“ + 1. That is, there exists some lower and upper part of the convex hull
point on these but below all other hyper - that arises from a Voronoi diagram, one
planes defined by S. Since polarity pre- obtains the closest site and furthest site
serves the relative position between Delaunay triangulation, respectively. In
points and hyperplanes, there is a hyper - such a triangulation, each edge is orthog-
plane with A(pl), . . . . A(p~) on it and onal to its dual facet. Conversely, the
A(q) above it, for all remaining sites existence of some set of sites allowing a

Voronoi Diagrams 9 385
triangulation whose edges are orthogo- rithms. For example, the planar Voronoi
nal to the facets of a given cell complex diagram of order-k can be obtained from
in R ~ is a necessary and sufficient condi- its order-(k – 1) predecessor in
tion for that cell complex to be a power O(kn log n) time by simply computing a
diagram or, equivalently, to be the pro- convex hull in R 3. Insertions and dele-
jection of a convex polyhedral surface tions of sites in an order-k Voronoi dia-
[Aurenhammer 1987b, 1987cI. If the gram also amount to computing certain
given cell complex is simple (i.e., exactly convex hulls; these operations even can
d + 1 cells meet at each vertex), a projec- be handled on-line in an efficient manner
tion surface can be reconstructed, if it [Aurenhammer and Schwarzkopf 19911.
exists, in time proportional to the num-
ber of facets of that cell complex. It is
3.1.4 k-Sets
well known that such a surface always
exists except for d = 2. This implies that Let Q be a set of n points in Rd+l. A
any simple complex of n cells in d > 3 k-set of Q is a subset of k points of Q
dimensions, which might have a large that can be separated from the remain-
number of lower dimensional faces, can ing n – k points by some hyperplane of
be specified by storing only n half-spaces Rd+ 1. The discussion in the foregoing
of Rd~~. section reveals a correspondence between
Surprisingly, even higher order power k-sets and order-k power diagrams.
diagrams are dual to some convex hull in By polarity, points in Q can be trans-
Rd+ 1 [Aurenhammer 1990a]. For any formed into hyperplanes; each point q e Q
subset T of S with k sites let A(T) bijectively corresponds to a weighted site
denote the point in Rd+ 1 whose ortho- p in R d such that A(p) = ~ and thus to
gonal projection onto Rd and whose the hyperplane T(p) in R +1. Now, let
(d + l)S’ coordinate are M be a k-set of Q, and consider a hyper -
plane h that separates M from Q – M.
Without loss of generality, M lies below
~p and ~’[pup- W(P)],
PET h. Equivalently, the pole of h lies below
all polar hyperplanes of the points in M
and above all polar hyperplanes of the
respectively. This definition conforms to
points in Q – M. Hence the intersection
the case k = 1 since A(T) = A(p) when
of the corresponding lower and upper
T = { p]. Using arguments similar as for
half-spaces of R ‘+ ~ bounded by these
k = 1, one may verify that the lower part
polar hyperplanes is nonempty. This in-
of the convex hull of the point-set
tersection, being a cell in the k-level of a
hyperplane arrangement, projects to a
Sk={ A(T)l TcS,l Tl=k} cell of the order-k power diagram of a set
of weighted sites, { p e Rd I A(p) c Q}.
is dual to k-PD(S). As one corollary, We have obtained a one-to-one corre-
that very lower part has to be dual to the spondence between k-sets in R d + 1 and
intersection of some corresponding upper power cells in R‘. This allows us to ap-
half-spaces and hence to an order-1 power ply known results on the number of k-sets
diagram. To be more precise, k-PD(S) to the analysis of the size of order-k power
coincides with PD( S&), for S& being the diagrams. Let fd( k, n) denote the maxi-
(accordingly weighted) orthogonal projec- mum number of k-sets of any set of n
tion of S’h onto R‘. This implies that the points in Rd. Known asymptotic values
class of power diagrams is closed under of this function are fz(k, n) = Q( n log k)
order-k modification— another result that and fz(k, n) = 0( n ~) [Edelsbrurmer
is not true for the more restrictive class and Welzl 1985], f~(k, n) = Q(nk log k)
of classical Voronoi diagrams. From the and fa(k, n) = 0(ns13 log5/3 n)
algorithmic viewpoint, k-PD(S) becomes [Edelsbrunner et al. 1986; Aronov et al.
constructible via convex hull algo - 1990, respectively]. Not surprisingly, a

better understood quantity is higher dimensions. Below we will show

that a certain rather general type of
Voronoi diagram in Rd - 1 can be embed-
ded into a power dia
end we identify RFy%h~~~~~~~~
Chazelle and Preparata [1986] showed sub;pace Xd = O of R‘. Let p be some
g~(k, n) = O(nk5) and Clarkson [19871 point site in R‘ - 1 and consider an arbi-
showed g~(k, n) = O(nl+’k2), for an~ trary distance function f( x, p) for p. For
e > 0. Note that g~(n – 1, n) = O(n ) some strictly increasing function F on
according to the maximum number of R, we define the cone of p with respect
cells of a hyperplane arrangement in to F as
Rd. Recently, Clarkson and Schor cone~(p)
[1989] succeeded in proving g~(k, n) =
@(n L~/zJ k FW21 ) as n/k ~ m for fixed —
— {(:@Rd-’,xd=F(f (x, p))}.
d. Determining exact or at least asymp-
totically tight bounds on ~d( k, n) still
A Voronoi diagram for point sites in
remains an important open problem.
R‘ - 1 under the distance function f is
termed transformable if there exists an
3. 1.5 Related Diagrams
F that forces the affine hull, U( p, q), of
The central role of power diagrams within cone~( p) n cone~( q) to be a hyperplane
the context of Voronoi diagrams becomes in R d for any distinct sites p and q.
even more apparent by the observations Transformable diagrams represent pro-
described below [Aurenhammer and Imai jected sections of power diagrams in R‘.
1988]. Let p, q, and r denote three A point x c Rd-~ satisfies f( x, p) <
weighted sites. For the three chordales f( X, q) if and only if its projection, x’,
they define, chor( p, q) fl chor( q, r) C onto cone~( p) lies in a fixed (open) half-
chor( p, r) necessarily holds. It is not dif- space, h(p, q), bounded by Q(P, q). As
ficult to see that this condition is also a consequence, x belongs to p’s region
sufficient for three hyperplanes to be the exactly if
chordales defined by three sites if we x’~z=
keep in mind the relationship between (1 ~(P,~),
qes – { p}
chordales in R d and intersections of hy -
perplanes in R d+ I . On the other hand, for S being the underlying set of sites in
the condition is trivially fulfilled by the Rd- 1. But the d-dimensional polyhedron
separators for any Voronoi diagram with Z is a power cell since a(p, q) n cr(q, r)
polyhedral regions since the regions c a( p, r) holds so that these affine hulls
would not form a cell complex otherwise. can be interpreted as chordales. In con-
We thus conclude that any Voronoi dia- clusion, the boundary of p’s region is the
orthogonal projection onto Rd-1 of
gram whose separators are hyperplanes
is a power diagram. This result applies cone~( p) intersected with a power cell in
to diagrams defined by the general Rd.
quadratic-form distance Transformable Voronoi diagrams occur
for the additively weighted distance,
Q(x, p) = (~-p) %(x-p) - ZO(p),
a(x, p) =~(x, p) – w(p)
with M a nonsingular and symmetric with F(a) = a,

(d x d)-matrix and thus to many partic-
ular distance functions considered in the and the multiplicatively weighted dis-
literature. Note that Q equals the power tance,
function if M is the identity matrix. 6(X, p)
One is tempted to believe that power m(x, p) = with F(m) = 2 m2.
diagrams are related solely to objects in w(p)

Voronoi Diagrams 0 387
_.-— — -— —-- . . A
----____
—.
y-— —B J-f-
“i
——— i —-——— —
\
N,
~,
‘+
\
\ “q
/
/ \, v
/
v
Figure 35. Cones for two distance functions.
The former gives rise to hyperbolically separator— and paraboloids of revolution

bounded and star-shaped regions (the in the latter case— intersecting in an el-
Johnson-Mehl model, Figure 6), whereas lipsis that projects down to a circular
the latter induces spherically bounded separator. In both cases, the affine hull
regions (the Apollonius model, Fig- of cone intersections is a plane.
ure 12). Its regions are disconnected, in Upper bounds on the size of both types
general, which already reflects the major of diagram in d dimensions can be de-
complexity of the latter type. rived from the maximum size of a power
Figure 35 displays cones for the addi- diagram in R ~+ 1. Concerning the algo-
tively weighted distance (above) and the rithmic aspects of this relationship, the
multiplicatively weighted distance (be- only known algorithms for constructing
low). They are cones of revolution in the these diagrams in d > 3 dimensions are
former case—intersecting in a hyper- obtained. The Apollonius model in the
bola that projects down to a hyperbolic plane can be computed in 0( n2) time and

388 e Franz Aurenhammer
space, which is worst-case optimal. We to a window, W, a segment on a line

mention that the notion of embeddability avoiding the convex hull of S. A site
generalizes nicely to order k. p G S and a point x are called visible if
the line segment joining x and p inter-
3.1.6 Upper Envelopes sects W. The peeper’s Voronoi diagram
of S and W assigns each point in the
Let fl, ..., f, be piecewise linear d- plane to the region of the closest site
variate functions on Rd. In the follow-
visible from it. See Figure 36 where the
ing, we will not distinguish between a
window is shown between two bold line
function on Rd and its graph in Rd+ 1. segments. The diagram is composed of
Consider the pointwise maximum, U, of
perpendicular bisectors of sites (dashed)
these functions: and of rays through sites and window
u(x) = mm, f,(x). endpoints (solid).
Each site p may be associated with an
U is called the upper envelope Of unbounded convex polygon plate(p) in
f,,... > L. We have already
observed that R3, being the projection onto T(p) of the
certain upper envelopes are related to set of all points visible from p. Again it
power diagrams. For any set S of is easy to see that a point x falls into the
weighted sites in R‘, the upper envelope region of a site p just if the vertical line
of the hyperplanes T(p), p ~ S, projects through x intersects plate(p) in a point
to PD(S). Below we briefly consider two that lies above plate(q) for all q e S –
types of Voronoi diagrams where this re- { p}. We conclude that the upper enve-
lationship generalizes particularly well. lope of plate( p), p e S, projects to the
Let S be a set of n point sites in the peeper’s Voronoi diagram of S.
plane, and consider a partition of S into Upper envelopes of piecewise linear
subsets Cl, . . . . C~ called clusters. The functions on R2 are well-studied geomet -
Hausdorff distance of a point x to the ric objects. Path and Sharir [1989] showed
cluster C, is defined as that they can attain a size of @(n2a(n)),
where n is the total number of triangles
needed in partitioning the linear pieces
h is also called the complete linkage dis- of these functions and where a denotes
tance; compare Section 2.5.1. Each clus- the inverse of Ackermann’s function.
ter C, may be associated with a convex Construction algorithms exist that run
polyhedral surface cap(C,) in R3, being in time proportional to their worst-case
the boundary of the polyhedron that size; see Edelsbrunner et al. [19891 who
comes from intersecting the half-spaces also give a lower bound of fl( n2 ) on the
below T(p) for all p e C,. Easy argu- size of a cluster Voronoi diagram, Re-
ments show the following. For any point cent results on upper envelopes by
x in the plane, h(x, C,) < h( x, C~) if and Huttenlocher et al. [1991] imply that the
only if the vertical line through x inter- diagram for t clusters with a total of n
sects cap(Ci) in a point lying above sites has a size of O(tna( tn)). The upper
cap(C7). This implies that the upper en- envelope arising from the peeper’s
velope of cap(C,), . . ., cap(CJ projects to Voronoi diagram has a worst case size of
the Voronoi diagram induced by 0( n2 ) [Aurenhammer and Stockl 1988].
cl, . . ., Ct and h. For an illustration of Note that the superlinear size of a
the Voronoi diagram for two-site clusters diagram implies the existence of
see Figure 30.13 disconnected regions.
To obtain another example, assume
that visibility of the sites is constrained 3.2 Topology of Planar Diagrams:
Divide-and-Conquer Construction and Its
13 Observe that the upper envelope of cap(T), for all Variants
subsets 2’ of k sites of S, consists of faces of the
k-level of {T(p) I p c S} and thus projects to the By far not all types of Voronoi diagrams
order-k Voronoi diagram of S. considered in the literature can be

\
\
\
\
\
\ ●
-- 9
----
Pa
-’\a\
.- —.- --—- ——.
———-
P
--., . .
0 b --._~
Figure 36. Peeper’s Voronoi diagram.
brought into connection with geometric distance function or the Minkowski dis-
objects having nice algorithmic proper- tance [Chew and Drysdale 1985]. Let C
ties, like convex hulls or hyperplane ar- denote the boundary of some compact and
rangements. So there is need for finding convex subset of R2 with the origin o in
features common to such “ungeometric” its interior. The distance with respect to
diagrams. We do this by extracting topo- C of o to some point x e R2 is given by
logical properties, being guided by their
relevance to the divide-and-conquer con- 6(X, o)
struction of Voronoi diagrams. Attention dc(x, O) =
fs(x’, o)
is restricted to the plane as the instance
of most applications and easiest analysis. for x’ being the point of intersection of C
The concept of diagram we are inter- and the ray from o to x. Clearly dc can
ested in here may be characterized as be defined with respect to any site q of a
follows. The n sites are points in R2. given set S by translating C so that q
The distance function varies by different occupies the former position of o, Ob -
shapes of its “circles” that may depend serve that dc( x, q) depends on the eu -
on local properties of R2. It thus may clidean distance ti( x, q) as well as on the
vary for the individual sites. direction of the ray from q to x. Convex
distance functions include, among others,
3.2.1 Convex Distance Functions the general LP-metric
An important type fitting into the con-
cept above is generated by the convex ~p(~,~)=vlql -x,lp+lqz-xglp,
*q
with
J*
xl
() ()
91
and x =
‘= 92 X2 ‘
whose most important instances are for

p = 1 (the Manhattan metric; C is a
square rotated by 450), p = 2 (the eu -
clidean distance; C is a circle), and p = co r
(the Maximum metric; C is an axis-

parallel square). Note that dc is a met-
ric only if C is point-symmetric
with respect to q since dc( q, r) might Figure 37. Redefining a separator.
differ from dc(r, q) otherwise. However,
the triangle inequality
That is, each point in R2 belongs to the
dc(q, r) + d~(r, s) ~ dc(q, s) region of the lexicographically least
can easily be shown to hold even if C is site among its closest sites in S. This
not point-symmetric. ensures that the interiors of the regions
Let us now investigate the Voronoi di- cover the plane up to a set of Lebesgue
agram defined by dc and a set S of n measure zero.
sites in R2. From the classical type we Defined in this way, the Voronoi dia-
adopt the notions of separator sep( q, r) of gram, VC(S), for a convex distance func-
q and r, dominance dom( q, r) of q over tion dc has the nice property that its
r, and region reg( q) of q; compare the regions are star shaped; Reg( q) contains
Introduction. Clearly sep( q, r) need not the straight-line segment between q and
be a straight line as is the case for x for each x e Reg( q), To prove this by
euclidean distance. An even more un- contradiction, assume the existence of a
pleasant phenomenon is that sep( q, r) point y on this segment with dc( y, q) >
fails to be one dimensional in general. To dc(q, r), for r ~ S – { q}. This implies
remedy this shortcoming, the notions dc(x, q) = dc(x, y) + dc(y, q) > dc(x,
above may be redefined slightly by refer- y) + dc( y, r). By the triangle inequality,
ring to the lexicographical order (sym- the latter sum is greater than or equal to
bolized by < ) of the sites in S [Klein dc(x, r) so that dc(x, q) > dc(x, r), that
and Wood 19881. The dominance of q is, x # Reg( q), follows.
over r is redefked as The star shapedness of the regions im-
plies their simple connectedness. Clearly,
Dom(q, r) regions cannot vanish since q ~ Reg( q)
holds for all q c S. So VC(5’) can be
dom(q, r) –sep(q, r), ifq<r viewed as a planar graph with exactly n
—
—
dom(q, r), otherwise. regions. Since vertex degrees are at least
{
three, there are 0(n) edges and vertices.
The new separator, Sep( q, r), is the in- We conclude a linear bound on the size of
tersection of the boundaries of Dom( q, r) Vc(s).
and Dom( r, q), which is obviously one-
dimensional. Figure 37 shows a two- 3.2.2 Divide-and-Conquer Construction
dimensional separator in the L1-metric
(left) and its redefined version (right). Our next aim is to show that Voronoi
For a Voronoi diagram in the L1-metric, diagrams for convex distance functions
see Figure 3. The region of a site q now can be constructed efficiently by a
can be rewritten as diuide-and-conquer algorithm. Divide
and conquer splits the problem at hand
Reg(q) = ,e~~{,} Dom(q, r). into two smaller subproblems, computes
their solutions recursively (unless they

are very small and can be solved by

trivial methods), and finally combines the \
\ e
partial solutions to the global one. The \
Vc type of diagram is well suited to at- \
\ ----
tack by this strategy. The underlying set --
“q
S of sites is partitioned into two subsets
e
S’l and S2 of nearly equal cardinal- e
r
ity. VC(SI) and VC(S2) are computed I
recursively and are then merged to I
+ I
Vc(s).
The merge step deserves the main at-
tention since it actually constructs the
diagram. We will discuss this process in Figure 38. Impossible merge chain
some detail and under the simplifying
but computationally unrestrictive as-
sumption that S1 and S2 are separable
by a vertical line. Since VC(SI) as well as The monotonicity of M(SI, S2) implies
VC(S.J defines a partition of R2, any its connectedness and unfoundedness.
point x ~ R2 falls into the region Regl(q) Hence the merge chain may be con-
of VC(SI) for some site q e S1 and into structed edge by edge, starting with an
the region Reg2(r) of VC(S2) for some unbounded one. As is mentioned in the
site rs S2. Now observe that x e Reg(q) Introduction, unbounded edges separate
of VC(S) if dc(x, q) < CZc(x, r), and x c sites on the convex hull of S. So it suf-
Reg(r) of VC(S) otherwise. So we have to fices to determine a site in S1 and a site
cut off some part from Regl( q) and from in S2 that are neighbors on the boundary
Reg2(r) by means of Sep(q, r) in order to of the convex hull of S. This can be
obtain Reg( q) and Reg( r). Carrying out accomplished in 0(n) time by finding a
this task for all relevant pairs of sites line tangent to the convex hulls of re-
constitutes the merge process. The union spective S1 and S2. Each further edge of
of the newly integrated pieces of separa- M(SI, S2) is found by tracing the bound-
tors will be called the merge chain, aries of the current regions of V(SI) and
M(Sl, S2). V(S2) in an appropriate direction until
The efficiency of an implementation of they intersect the actual separator. Easy
the merge process critically depends on counting arguments show that only 0(n)
the topological properties of the merge edges are processed in total: Each edge
chain. We first show that M(SI, S2 ) be- traced is either discarded or shortened to
haves well if the euclidean distance 8 is a new edge of V(S). This implies an
taken for dc [Shames and Hoey 1975]. overall runtime of 0(n) for the merge
As a matter of fact, each horizontal line process. By the recurrence relation
intersects M(SI, S2 ) in exactly one point l“(n) = 2T(n /2) + O(n) that results
in this case, that is to say, M(SI, S2) is from using divide and conquer, the
(vertically) monotone. M(SI, S2) is a time complexity T(n) for computing a
polygonal line composed of edges of the classical Voronoi diagram is 0( n log n).
classical diagram V(S) that are portions Let us now come back to the distance
of perpendicular bisectors of sites in S1 dc for general convex shapes C [Chew
and S2, respectively. Figure 17 gives an and Drysdale 1985]. As in the euclidean
illustration. If there were a horizontal case, M(SI, S2 ) is composed of those parts
line intersecting the merge chain in two of the boundaries of regions of VC(S)
or more points, some site q ● SI would be that separate sites in S1 from sites in S2.
forced to have a larger xl-coordinate than Thus M(SI, S2) is one dimensional. Po-
a site r~S2. This contradicts the as- tentially, this merge chain may consist of
sumption that S’l and S2 are separable cyclic and acyclic (but then necessarily
by a vertical line; compare Figure 38. unbounded) topological curves in R2.

Examples show that M(SI, SQ) is in fact

disconnected in general. Its connected
components all, however, are acyclic and
thus unbounded. Let us prove that there
are no cycles by assuming the existence
of some cycle M and deducing a contra-
diction. Without loss of generality, M /
encircles a subset T of sites in SI.
/
Clearly, then M does not encircle any
/
site in Sz. Furthermore, there exists a
/ 0,
leftmost point x e M with x e Sep(r, s)
for some r, s e Sz. Since the regions of /
VC(S) are star shaped, the three open I
line segments from x to r, s, and some \x
q e T, respectively, do not intersect M.
Thus, as r and s lie to the right of q, x \
has to lie to the left of q. This, however, \
contradicts the convexity of the underly- \
ing shape. C encircles x and passes \
through q, r, and s since the last three \
points are equidistant from x with re-
spect to dc. Figure 39 illustrates this
situation; M and C are shown bold and
dashed, respectively.
In order to construct the acyclic merge
chain M(S1, S2), one needs to detect some
unbounded edge of each of its connected
components. The previously described ,
procedure of tracing boundaries of re-
gions may then be applied. Unfortu-
nately, the convex-hull method that Figure 39. Cycles do not occur.
works successfully in the euclidean case

fails to be correct for dc. To get an alter-
native method, we observe that each un- x~R2, and, in addition, its separators
bounded edge has to be contained in the can be computed and intersected in 0(1)
intersection of two unbounded regions of time. This may not be the case for
VC(SI) and VC(S2), respectively. By a complex shapes C, however.
circular scan through these regions, each
relevant pair Regl( q) and Regz(r) can be
3.2.3 Nice Metrics
found and intersected with Sep( q, r) in a
total of 0(n) time. If Regl(q) fl Regz(r) We have seen that divide and conquer is
n Sep( q, r) extends to infinity, then this a powerful approach to the construction
edge is taken as the starting edge for a of generalized Voronoi diagrams. It also
connected component of M(SI, S2). applies well to sites more general than
The linear-time behavior of the merge points; see Section 1.3.3. We do not pur-
process thus can be maintained for gen- sue these modifications here, however,
eral convex distance functions dc. Con- since they do not fit into the present
sequently, VC(S) requires O(n log n) concept of a diagram. Instead, we at-
time and 0(n) space for construction. Of tempt to characterize metrics in Rz that
course, the complexity analysis is based preserve the two main properties of a
on the assumption that dc is computa- Voronoi diagram exploited in the pre-
tionally simple. That is, dc( q, x) can be ceding discussion [Klein 1989; Klein and
calculated in O(1) time for all q e S and Wood 1988]: the connectedness of regions

that implies an 0(n) size and the ab- a d-straight curve (dashed) between two
sence of cycles in the merge curve that points x and y and a Voronoi diagram
allows an 0( n log n) time construction. (solid) for eight sites in this metric. For
Let d be any metric in R2 that in- the geodesic distance among polygonal
duces the euclidean topology. A curve ~ obstacles, d-straight curves are the mini-
in R2 is termed d-straight if, for any mum euclidean length connections avoid-
three consecutive points x, y, z on -y, the ing any edge of the polygons (compare
equality Figure 26). Also of interest are composite
metrics where, for instance, distances are
d(x, y) + d(~, Z) = d(X, Z)
measured in the L1-metric in some por-
holds. Clearly &straight curves are tion of R2 and in the L2- (i.e., euclidean)
metric in the complement. Here, d-
straight-line segments, for ~ being the
straight curves are composed of those in
euclidean metric. By arguments similar
LI and L2 such that curve length is min-
to that used for convex distance functions
imized. In all three examples, the respec-
(namely using the triangle inequality), it
tive disks are simply connected. As a
can be verified that any Voronoi region
consequence, the resulting Voronoi dia-
Reg( p) for d is d-star-shaped: y c Reg( p)
holds for the d-straight curve -y from p grams-except for the geodesic distance
to any x e Reg( p). The d-star-shapedness
function that fails to be computationally
of Reg( p) does not necessarily imply its simple— can be computed in optimal
connectedness since d-straight curves do 0( n log n) time and 0(n) space by divide
and conquer. Thereby, the underlying set
not always exist. If, however, for any two
of point sites has to be divided into sub-
points x, z e R2 there is some y distinct
sets by curves whose intersections with
from x, z with d(x, y) + d(y, z) =
d( x, z), then the existence of -y is guar-
any disk for d is either connected or
empty. Consult Klein [1989] for further
anteed. Hence the Voronoi regions under
details. We pose it as an open question to
those metrics are connected.
find equivalent characterizations for
Posing an additional restriction on the
metrics whose Voronoi diagrams are
metric d makes the Voronoi diagram
even more well behaved: Assume that, constructible by divide and conquer in
for any point m e R2 and for any positive optimal time.
number r, the generalized disk
3.3 Deformation of the Voronoi Diagram:
D={xc R21d(x, m) <r} Plane-Sweep Technique
with center m and radius r under the The variety of techniques that can be
metric d is a simply connected set. For used for constructing Voronoi diagrams
fixed r, the shape of D may vary with is possibly as fascinating as the diagrams
the position of m provided D remains themselves. The construction of the
simply connected. Symmetric convex dis- classical planar Voronoi diagram by
tance functions dc (i.e., with point- means of divide and conquer or via
symmetric shape C) constitute a special three-dimensional convex hulls are
case of translation-invariant disks. And, well-investigated algorithmic problems.
indeed, the proof that merge chains are Although the achieved complexity bounds
acyclic can be extended from convex dis- are worst-case optimal, researchers con-
tance functions to d. tinued considering alternative methods
To underline the usefulness of this con- of construction. Aside from the incre-
cept, let us give some examples for d. Let mental insertion strategy described in
K be any point in R2. The Moscow met- Sections 1.3.1 and 1.3.2, the plane-sweep
ric with respect to K induces d-straight technique to be discussed now stands out
curves composed of pieces straight to- by its conceptual and computational sim-
ward K or radially around K of minimal plicity. It extends nicely to various
total euclidean length. Figure 40 depicts generalized types of Voronoi diagrams.

394 ●
Franz Aurenhammer
‘~
@
e
‘)/
Figure 40. Voronoi diagram in the Moscow metric [Klein 1989].
3.3.1 Deformation p c S} as the pointwise minimum of these

functions or, equivalently, as the surface
Let S denote a set of n point sites in R2
composed of that portion of each cone
and let V(S) be their Voronoi diagram.
that lies below all other cones. From the
As we will see, there are serious reasons
definition of cone(p) it is evident that
for refraining from a direct construction
this lower envelope projects vertically
of V(S) via the plane-sweep technique.
to V(S) onto R2.
Therefore, first we introduce a continu-
For our purposes it is preferable to
ous deformation of V(S) [Fortune 19871.
project the envelope onto R2 in a differ-
Then, we describe the plane-sweep tech-
ent way, namely under an angle of T/4
nique and its use for computing this
in positive xz-direction. This already
deformation.
gives the desired deformation V*(S) of
To explain the mechanism of deforma-
V(S). To be more precise, the deforma-
tion, the euclidean distance function with
tion * maps each point x e reg( p), say
respect to the sites is interpreted in the
following way. With each site p G S, a
(
xl
()
xl
cone x= into x* =
X2 ‘ x2+a(x, p) ) “
cone(p)
Let us study what happens in this pro-
={(X, Z) GR31X~R2, Z=8(X, p)} cess. Clearly, sites are invariant under *.
Each separator sep( p, q) is deformed into
is associated. Cone(p) is an upwardly di-
a hyperbola with bottommost point p if
rected cone in R 3 with vertical axis of
p is below q and bottommost point q,
revolution, with apex p, and with an
otherwise. (This hyperbola degenerates
interior angle of m/2.14 Cones may be
to a vertical half line if p and q have the
viewed as bivariate functions on R2. We
same xz-coordinate.) Consequently, the
define the lower envelope of {cone(p) I
deformed region reg*( p) of p is the in-
tersection of hyperbolically bounded half
planes of R2. Reference to the upper en-
14Note that the definition of cone(p) agrees with
velope model shows that * preserves the
that of coneF( p) used for transformable Voronoi
diagrams in Section 3.1.5; F is the identity topological properties of a Voronoi dia-
function. gram. For instance, the interiors of the

\*,,f
. .,. \ I
Initially,
tion of V*(S)
Voronoi Diagrams
when L is below S, the por-

below L contains no de-
“ 395
+’
,.’ formed edge. Hence the cross section is
. . .. ,/”, /) ---
-.*. ‘. *..
empty. As L moves upward the cross
. .
‘.\ section has to be updated, either by start-
. . ~__-—-
.
ing a new region when L hits a site or by
starting a new edge when L hits a vertex
that comes from intersecting two de-
formed separators. The properties of
Figure 41. Deforming a Voronoi diagram [Fortune
1987].
V*(S) described above will ensure the
correctness of this method. Efficiency is
gained by organizing points of intersec-
tion of L with edges of V*(S) (the cross
deformed regions cover R2 up to a set of section) in a dictionary ordered by xl-
Lebesgue measure zero. The algorithmic coordinates and by organizing sites and
advantage of the deformation over the already detected vertices (the critical
original is that the bottommost point of points) in a priority queue ordered by
each region is its defining site. Excep- xz-coordinates. There are at most n +
tions are regions for the bottommost sites (2 n - 4) critical points since V(S), and
in S that necessarily are unbounded. thus V*(S), has at most 2 n – 4 vertices;
Figure 41 shows the Voronoi diagram of compare the Introduction. Further, there
six sites (left) and its deformation (right). are at most n – 1 points in the cross
section since L cannot intersect a region
3.3.2 Plane-Sweep Technique
of V*(S) twice. An update causes only a
constant number of operations on both
The properties of V*(S) are now ex- data structures each of which cost
ploited for its construction using the O(log n) time. In conclusion, V*(S) is
plane-sweep technique. Generally, this computed in 0( n log n) time and 0(n)
technique proceeds as follows. A horizon- space. It is clear that the original
tal line L is swept across the object to be diagram V(S) can be derived from its
constructed from below by keeping the deformation in 0(n) time.
invariant that the portion of the object Fortunately, this simple construction
below L is complete at any point in time. method extends to several generalized
During the plane sweep, the cross section Voronoi diagrams [Fortune 1987]. It nat -
of L with this object has to be updated at urally applies to the additively weighted
certain critical points. We thus have to type since the correspondence to lower
handle a one-dimensional dynamic prob- envelopes extends provided cone(p) is
lem instead of a two-dimensional static translated downward by the weight W(p)
problem. for each site p. The intersection of
The applicability of the plane-sweep cone(p) with R2 is then a circle with
technique crucially depends on whether center p and radius w ( p), and indeed
the “critical points” can be predicted. this type may be interpreted as the
This is, for example, not the case if the Voronoi diagram for circles under the
Voronoi diagram of S is to be constructed euclidean distance function (Johnson-
in its original: A site p e S might lie Mehl model, Figures 6 and 35). The time
above L while its region reg( p) already bound of 0( n log n) and the space bound
extends to below L. And it is important of 0(n) clearly remains unaffected.
to know when regions begin to intersect An important type suited for a plane-
L in order to integrate them properly sweep attack is the Voronoi diagram for
into the cross section. We do not run into noncrossing straight-line segments in R2.
such problems when constructing V*(S), The distance of a point x from a segment
however, since the bottommost point of a s is measured by min{~(x, y) I y e S}. The
deformed region is just its defining site. region of a segment can be deformed so

as to have the lower endpoint of that regions it supports. Let t support

segment as its bottommost point. Al- ‘egb(pl), . . . . reg~(p~) frOm the left and
though details are more complicated, a reg~(ql), ..., reg ~(ql) from the right. The
reasonably simple 0( n log n) time and former b-regions are extended to the left
0(n) space algorithm is obtained. In par- of t as if only sites pl, . . ., ph were pres-
ticular, the medial axis of a simple poly- ent, and the latter b-regions are ex -
gon is constructible via plane-sweep: The tended to the right of t as if only sites
edges of the polygon are taken as line ql,..> ql were present, Note that p, = qj
segment sites, and the parts of the dia- has to occur twice, namely, for the end-
gram outside of the polygon are ignored. points r and s of t. So reg~(r) and reg~(s)
Note that the separator of two segments are extended to both sides of t.The c-re-
consists of portions of straight lines and gion, reg.(p), of a site p e S is defined as
parabolas; compare Figure 25. the union of reg ~(p) and its extensions
with respect to all obstacles. c-regions
3.3.3 Constrained Voronoi Diagrams clearly may overlap, so they fail to define
a cell complex and the graph defined by
A further type of Voronoi diagram, and their edges is nonplanar. See Figure 42
one to which the plane-sweep technique illustrating the bounded Voronoi dia-
applies particularly nicely, is the classi- gram (solid) and the constrained Voronoi
cal diagram constrained by a set of line diagram (solid or dashed or dotted) for 11
segments [Seidel 1988]. Let S denote a sites and 2 obstacles (bold).
set of n point sites in R2 and let T be a The significance of VC(S, T) is due
set of m noncrossing line segments with to its dual structure, the constrained
their endpoints in S. Note that m = O(n) Delaunay triangulation of S and T.
since the segments form a planar graph. Intuitively, this triangulation is as simi-
We view the segments in T as obstacles lar as possible to the classical Delaunay
and define the bounded distance between triangulation but integrates the obsta-
two points x, y ~ R2 as cles in T, a structure with several practi -
b(x, y) cal applications. Formally, it consists of
the obstacle segments in T and, in addi-
— a(x, y), if~fl T.Ql, tion, of all edges between sites p, q e S
that have b( p, q) < m and that lie on a
{ co, otherwise.
circle enclosing only sites r ~ S with at
The b-regions least one of b(r, p), b(r, q) = m. The in-
terested reader may verify that two sites
reg~(~) are connected in this triangulation if and
={xe R21b(x, p)= b(x, q), qeS} only if their perpendicular bisector con-
tributes an edge to the constrained
for all sites p ● S define the bounded Voronoi diagram. Note that the bisector
Voronoi diagram, V~(S, T), of S and T. of the endpoint sites of an obstacle may
b-regions of endpoint sites are generally contribute two edges (on either side of
nonconvex (near the corresponding ob- the obstacle) rather than only one. Since
stacle), whereas b-regions of nonendpoint S contains n sites, T contains m obsta-
sites have to be convex. Both types of cles, and no triangulation of n points can
b-regions, and necessarily the former have more than 3 n – 6 edges, we
type, might have portions of obstacles conclude that VC(S, T) realizes at most
rather than of perpendicular bisectors as 3 n + m – 6 edges and thus has a size of
edges. Clearly, V~(S, T) is a polygonal 0(n). Figure 27 shows the constrained
cell complex in R2. Delaunay triangulation that corresponds
The bounded Voronoi diagram is now to the constrained Voronoi diagram in
modified to the constrained Voronoi Figure 42.
diagram, VC(S, T), of S and T by modi- Let us now come to the plane-sweep
fying, for each obstacle t ~ T, the b- construction of VC(S, T). To aid the

points arising during the construction of

I V$(S, and all V,* is proportional
T) to
the size of VC(S, T) and thus is 0(n).
Using appropriate data structures sup-
porting the m + 1 plane-sweeps allows
..” us to implement the operations at each
critical point in O(log n) time so an
0( n log n) time and 0(n) space algo-
. rithm for computing VC(S, T) is obtained.
In particular, the constrained Delaunay
triangulation of S and T is computable
within these bounds since it can be ob -
/ tained in O(n) time from VC(S, T) by
exploiting duality.
Figure 42. Bounded and constrained Voronoi dia-
gram [Seidel 1988]. ACKNOWLEDGMENTS
This survey was completed while the author was at

the Institut fur Informatik, Freie Universitat
intuition, we think of VC(S, T) as being Berlin, Germany. Work was partially supported by
embedded into m + 1 parallel planes in the ESPRIT II Basic Research Program of the EC
the following way. One plane, call it Ho, under contract no. 3075 (project ALCOM).
contains the bounded Voronoi diagram

REFERENCES
V~(S, T). Each of the m planes, 11~, for
the obstacle t e T contains the diagram, AGARWAL, P. K., EDELSBRUNNER, H., SCHWARZKOPF,
Vt, induced by the extensions with re- O., AND WELZL, E, 1990. Euclidean minimum
spanning trees and bicromatic closest pairs. In
spect to t for all b-regions supported by Proceedings of the 6th Annual ACM Sympo-
t.Clearly, this embedding comprises the sium on Computational Geometry, pp. 203-210,
same information as is inherent in AGGARWAL, A., AND SURI, S. 1987. Fast algo-
VC(S, T). The idea now is to sweep HO rithms for computing the largest empty rectan-
and all Ht simultaneously by (essen- gle. In Proceedings of the 3rd Annual ACM
tially one and the same) horizontal line Symposium on Computational Geometry, pp.
278-290.
L in order to construct Vb(S, T) and all
AGGARWAL, A., HANSEN, M., ANDLEIGHTON, T. 1990.
Vt, for tE T, separately. Again these m Solving query retrieval problems by compact-
+ 1 diagrams are not constructed di- ing Voronoi diagrams. In Proceedings of the
rectly but via their deformations 22nd Annual ACM Symposium on STOC, pp.
331-340.
V:(S, T) and V~* that come from map-
ping upward, within the respective plane, AGGARWAL, A., GUIBAS, L. J., SAXE, J., AND SHOR,
P. W. 1989a. A linear time algorithm for
each point ~ inside an (extension of a)
computing the Voronoi diagram of a convex
b-region reg~( p) by 8(x, p). Since an ex- polygon. Discrete Comput. Geometry 4,
tended b-region is just a classical Voronoi 591-604.
region, its deformation has the defining AGGARWAL, A., IMAI, H., KATOH, N., AND SURL S.
site as the bottommost point (unless the 1989b. Finding k points with minimum di-
ameter and related problems. In Proceedings of
site itself is bottommost for the respec-
the 5th Annual ACM Symposium on Computa-
tive plane). In analogy to the classical tional Geometry, pp. 283-291.
Voronoi diagram, regions are created as AGGARWAL,A., CHAZELLE,13., GUIBAS, L. J.,
sites are hit by L, and edges are created O’DUNLAING, C., AND YAP, C. K. 1988. Paral-
as L hits intersections of deformed bisec- lel computational geometry. Algorithmic 3,
tors. One must, however, take care that 293-327.
edges are assigned correctly to their em- AHUJA, N. 1982. Dot pattern processing using
Voronoi polygons as neighborhoods. IEEE
bedding planes. When a deformed edge Trans. Putt. Anal. Mach. Int. PAMI-4, 336-343.
in Ho meets a deformed obstacle t*,then AKL, S. 1983 A note on Euclidean matchings, tri-
its construction is continued within Ht angulations, and spanning trees. J. Comb. Infi
and vice versa. The number of critical Syst. Sci. 8, 169-174.

ALEXANDERSON, G. L., AND WETZEL, J. E. 1978. cell complexes and constructing projection
Simple partition of space. Math. Magazine 51, polyhedra. J. Symbolic Comput. 3, 249-255.
220-225. AURENHAMMER, F. 1988a. Improved algorithms for
ALT, H., AND YAP, C. K. 1990. Algorithmic as- discs and balls using power diagrams. J. Algo-
pects of motion planning A tutorial (part 2). rithms 9, 151-161.
Algorithms Rev. 1,61-77. AURENEIAM~ER, F. 1988b. Linear combinations
AONUMA, H., IMAI, H., IMAI, K., AND TOKUYAMA, T. from power domains. Geometriae Dedicata 28,
1990. Maximin location ofconvexobjects ina 45-52.
polygon and related dynamic Voronoi dia- AURENHAMMER, F. 1990a. A new duality result
grams. In Proceedings of 6th Annual ACM concerning Voronoi diagrams. Discrete Com-
Symposium on Computational Geometry. put. Geometry 5, 243-254.
pp. 225-234.
AURENHAMMER, F, 1990b. A relationship between
ARNOLD, D. B., AND MILNE, W. J. 1984. The use of
Gale transforms and Voronoi diagrams. Dis-
Voronoi tessellations in processing soil survey
crete Appl. Math. 28, 83-91.
results. IEEE Comput. Graph. Appl. 4, 22-30.
AUEENIiA~~ER, F., AND EDELSBRUNNER, H. 1984.
ARONOV, B. 1989. On the geodesic Voronoi dia-
An optimal algorithm for constructing the
gram of point sites in a simple polygon. Algo-
weighted Voronoi diagram in the plane.
rithmic 4, 109–140.
Pattern Recognition 17, 251-257.
ARONOV, B., FORTUNE, S., AND WILFONG, G. 1988.
AURENHAMMER, F., AND IMAI, H. 1988. Geometric
The furthest-site geodesic Voronoi diagram. In
relations among Voronoi diagrams. Geometriae
Proceedings of the 4th Annual ACM Sympo-
Dedicata 27, 65-75.
sium on Computational Geometry. pp. 229-240.
AUKENHAMMER, F., AND SCHWARZKOPF, O. [1991].
ARONOV, B., CHAZELLE, B., EDELSBRUNNER, H.,
A simple on-line randomized incremental algo-
GUIBAS, L. J., SHARIR, M., AND WENGER, R.
rithm for computing higher-order Voronoi dia-
1990. Points and triangles in the plane and
grams. In Proceedings of the 7th Annual ACM
halving planes in space. In Proceedings of the
Symposium on Computational Geometry. pp.
6th Annual ACM Symposium on Computa-
142-151.
tional Geometry. pp. 112-115.
AURENHAMMER, F., AND STOCKL, G. 1988. On the
ASANO, T., AND ASANO, T. 1986. Voronoi diagram
peeper’s Voronoi diagram. Rep. 264, IIG-TU
for points in a simple polygon. In Perspectives
Graz, Austria.
in Computing 15, D. S. Johnson, T. Nishizeki,
A. Nozaki, H. S. Wilf, Eds. Academic Press, AURENHANtMER, F., STOCKL, G., AND WELZL, E. 1991.
New York, pp. 51-64. The post-ofllce problem for fuzzy point sets,
Rep. 91-07, FU Berlin, Germany.
ASAFJO, T., BHATTACHARYA, B., KEIL, M., AND
YAO, F. 1988. Clustering algorithms based on AVIS, D., AND BHATTACHARYA, B. K. 1983. Algo-
minimum and maximum spanning trees. In rithms for computing d-dimensional Voronoi
Proceedings of the 4th Annual ACM Sympo- diagrams and their duals. Adu. Comput. Res.
sium on Computational Geometry, pp. 252-257. 1, 159-180.
ASH, P. F., AND BOLKER, E. D. 1985. Recognizing BABENKO, V. F. 1977. On the optimal cubature
Dirichlet tessellations. Geometriae Dedicata 19, formulae on certain classes of continuous func-
175-206. tions. Analysw Math. 3, 3-9.
ASH, P. F., AND BOLKER, E. D. 1986. Generalized BADDELEY, A. 1977. A fourth note on recent re-
Dirichlet tessellations. GeometrLae Dedicata 20, search in geometrical probability. Adu. Appt.
209-243. Probab. 9, 824-860.
ASH, P. F., BOLKER, E. D., CRAPO, H., AND WHITE- BALTSAN, A., AND SHARIR, M. 1988. On shortest
LEY, W. 1988. Convex polyhedra, Dirichlet paths between two convex polyhedra. J. ACM
tessellations, and spider webs. In Shaping 35, 267-287.
Space: A Polyhedral Approach, M. Senechal BENTLEY, J. L., AND MAURER, H. A. 1979. A note
and G. Fleck, Eds. Birkhauser, Boston, pp. on Euclidean near neighboring searching in
231-250. the plane. Infi Process. Lett. 8, 133-136.
AUGENBAUM, J. M., AND PESKIN, C. S. 1985. On BENTLEY, J. L., WEIDE, B. W., AND YAO, A. C.
the construction of the Voronoi mesh on a 1980. Optimal expected-time algorithms for
sphere. J. Comput. Phys. 59, 177-192. closest-point problems. ACM Trans. Math.
AUEEN~A~MER, F. 1987a. Power diagrams: Prop- Sof%w. 6, 563-580.
erties, algorithms, and applications. SIAM J. BERN, M. W., EPPSTEIN, D., AND GILBERT, J. 1990.
Comput. 16, 78-96. Provably good mesh generation. Manuscript,
AURENHAMMER, F. 1987b. A criterion for the aftine XEROX Palo Alto Research Center, Palo Alto,
equivalence of cell complexes in R d and convex Calif.
polyhedra in R d+ ~, Discrete Comput. Geometry
BESAG, J. 1974. Spatial interaction and the statis-
2, 49-64. tical analysis of lattice systems. J. Roy. Statist.
AUREI-JHAIWER, F. 1987c. Recognizing polytopical SOC. B 36, 192-236.

BIEBERBACH, L. 1912. fier die Bewegnngsgrup- improved algorithm for constructing k th-order
pen der euldidischen Raume I, II. Math.Ann. Voronoi diagrams. IEEE Trans. Comput. C-36,
72,400-412. 1349-1354.
BLUM, H. 1967. A transformation for extracting CHAZELLE, B., AND PREPARATA, F. P. 1986. Half-
new descriptors of shape. In Proceedings of the space range search: an algorithmic application
Symposium on Models for the Perception of of k-sets. Discrete Comput. Geom. 1, 83-94.
Speech and Visual Form. Weiant Whaters- CHAZELLE, B., DRYSDALE, R. L., LEE, D. T. 1986a.
Dunn, Ed. MIT Press, Cambridge, Mass., pp. Computing the largest empty rectangle. SIAM
362-380. J. Comput. 15, 300-315.
BLUM,H. 1973. Biological shape and visual sci- CHAZELLE, B., COLE, R., PREPARATA, F. P., AND
ence(Part 1). J. Theor. Biol. 38, 205-287. YAP, C. K. 1986b. New upper bounds for
BOE+SONNAT, J.-D. 1988. Shape recognition from neighbor searching. Znf Control 68, 105-124.
planar cross sections. Comput. Vision Graph. CHAZELLE, B., EDELSBRUNNER, H., GUIBAS, L. J.,
and Image Process. 44, 1–29.
HERSHBERGER, J. E., SEIDEL, R., AND SHARIR,
BOISSONNAT, J.-D., AND TEILLAUD, M. 1986. A hi- M. 1990. Slimming down by addin~ selecting
erarchical representation of objects: The Delau- heavily covered points. In Proceedings of the
nay tree. In Proceedings of 2nd Annual ACM 6th Annual ACM Symposium on Computa-
Symposium on Computational Geometry, tional Geometry, pp. 116-127.
pp. 260-268.
CHERITON, D., AND TARJAN, R. E. 1976. Finding
BooTs, B.N, 1979. Weighting Thiessen polygons. minimum spanning trees. SIAM J. Comput. 5,
Econ. Geography, 248-259. 724-742.
BOOTS, B. N. 1986. Voronoi (Thiessen) polygons. CHEW, L. P. 1986. There is a planar graph almost
In CATit40G 45, Geo Books, Norwich, Corm. as good as the complete graph. In Proceedings
BOWYER, A. 1981. Computing Dirichlet tessella- of the 2nd Annual ACM Symposium on Compu-
tions. Computer J. 24,162-166. tational Geometry. pp. 169-177.
BRASSEL, K, E., AND REIF, D. 1979. A procedure to CHEW, L. P. 1989a. Constrained Delaunay trian-
generate Thiessen polygons. Geograph. Anal. gulations. Algorithmic 4, 97-108.
11, 289-303. CHEW, L. P. 1989b. Guaranteed-quality triangu-
BRISSON, E. 1989. Representing geometric struc- lar meshes. Rep. TR 89-983, Dept. Comput.
turesin dimensions: Topology and order. In Sci., Cornell Univ., Ithaca, N.Y.
Proceedings of the 5th Annual ACM Sympo-
CHEW, L. P., AND DRYSDALE, R. L. 1985. Voronoi
sium on Computational Geometry. pp. 218-227.
diagrams based on convex distance functions.
BRONDSTED, A. 1983. An Introduction to Convex In Proceedings of the 1st Annual ACM Sympo-
Polytopes. Springer, New York, Heidelberg, sium on Computational Geometry. pp. 235-244.
Berlin.
CHEW, L. P., AND KEDEM, K. 1989. Placing the
BROSTOW, W., AND SICOTTE, Y. 1975. Coordination largest similar copy of a convex polygon among
number in liquid argon. Physics A 80, polygonal obstacles. In Proceedings of the 5th
513-522. Annual ACM Symposium on Computational
BROSTOW, W., DUSSAULT, J. P., AND Fox, B. L. Geometry. pp. 167-174.
1978. Construction of Voronoi polyhedra. CHOW, A. 1980. Parallel algorithms for geometric
J. Comput. Phys. 29, 81-92. problems. Ph.D. dissertation. Dept. Comput.
BROWN, K. Q. 1979. Voronoi diagrams from con- Sci., Univ. of Illinois, Urbana, Ill.
vex hulls. Znfi Process. Lett. 9, 223-228. CHRISTOFIDES, N. 1976. Worst-case analysis of a
BROWN, K. Q. 1980. Geometric transforms for fast new heuristic for the traveling salesman prob-
geometric algorithms. Ph.D. dissertation, lem. Symposium on Algorithms and Complex-
Carnegie-Mellon Univ., Pittsburgh, Penn. ity, Carnegie-Mellon Univ., Penn.
BRUMBERGER, H. AND GOODISMAN. 1983. Voronoi CLARKSON, K. L. 1987. New applications of ran-
cells: An interesting and potentially useful cell dom sampling to computational geometry.
model for interpreting the small angle scatter - Discrete Comput. Geom. 2, 195-222.
ing of catalysts. J. Appl. Crystallogr. 16,83-88. CLARKSON, K. L. 1989. An algorithm for geomet-
CALABI, L., AND HARTNETT, W .E. 1968. Shape ric minimum spanning trees requiring nearly
recognition, prairie tires, convex deficiencies, linear expected time. Algorithmic 4,461-468.
and skeletons. Am. Math. Monthly 75,
CLARKSON, K. L., AND SHOR, P. W. 1988. Algo-
335-342.
rithms for diametral pairs and convex hulls
CANNY, J., AND DONALD, B. 1988. Simplified that are optimal, randomized, and incremen-
Voronoi diagrams. Discrete Comput. Geom. 3, tal. In Proceedings of the 4th Annual ACM
219-236. Symposium on Computational Geometry.
CHAZELLE, B. 1985. How to search in history. Inf pp. 12-17.
Control 64, 77-99. CLARKSON, K. L., AND SHOR, P. W. 1989. Applica-
CHAZELLE, B., AND Emmsmmmmm, H. 1987. An tions of random sampling in computational

400 e Franz Aurenhammer
geometry, II. Discrete Comput. Geom. 4, DILLENCOURT, M. B. 1987c. Toughness and Delau-
387-421. nay triangulations. In Proceedings of the 3rd
COLE, R., GOODRICH, M. T., AND O’DUNLAING, C. Annual ACM Symposium on Computational
1990. Merging free trees in parallel for effi- Geometry. pp. 186-194.
cient Voronoi diagram construction. Springer DIBICHLET, G. L. 1850. Uber die Reduction der
LNCS 443, pp. 432-445. positiven quadratischen Formen mit drei
CONWAY, J. M., AND SLOANE, N. J. A. 1982. unbestimmten ganzen Zahlen. J. Reme u.
Voronoi regions of lattices, second moments of Angew. Math. 40, 209-227,
polytopes, and quantization. IEEE Trans. Inf DOBKIN, D. P., AND LASZLO, M. J. 1989. Primi.
Theory IT-28, 211-226. tives for the manipulation of three-dimensional
CRAIN, L. K. 1978. The Monte Carlo generation of subdivisions. Algorithmic 4, 3-32.
random polygons. Comput. Geosci. 4, 131-141. DOBKIN, D. P., AND LIPTON, R. J. 1976. Multidi-
CRAPO, H. 1979, Structural rigidity, Struct. Topol. mensional searching problems. SIAM J. Com-
1, 13-45. put. 5, 181-186.
CRUZ ORIVE, L.-M. 1979. Distortion of certain DOBKIN, D. P., FRIEDMAN, S., AND SUPOWIT, K.
Voronoi tessellations when one particle moves. 1990. Delaunay graphs are almost as good as
J. Appl. Probab. 16, 95-103. complete graphs. Discrete Comput. Geom. 5,
399-407.
DAVID, E. E., AND DAVID, C. W. 1982. Voronoi
polyhedra as a tool for studying solvation DRYSDALE, R. L. 1990. A practical algorithm for
structure. J. Chem. Phys. 76, 4611 –4614. computing the Delaunay triangulation for con-
vex distance functions. In Proceedings of the
DE FLORIANI, L., FALCIDIENO, B,, PIENOVI, C., AND
1st Annual ACM-SIAM Symposium on Discrete
NAGY, G. 1988. On sorting triangles in a De-
Algorithms. pp. 159-168.
launay tessellation. Rep. Inst. Mat. Appl.,
Consiglio Nazionale dells Richerche, Geneva, DWYER, R. A. 1987. A faster divide-and-conquer
Italy. algorithm for constructing Delaunay triangula-
tions. Algorithmic 2, 137-151.
DEHNE, F., AND KLEIN, R. 1987. An optimal algo-
rithm for computing the Voronoi diagram on a DWYER, R. A. 1989. Higher-dimensional Voronoi
cone. Rep. SCS-TR-122, School Comput. Sci., diagrams in linear expected time. In Proceed-
Carleton Univ., Ottawa, Canada. ings of the 5th Annual ACM Symposium on
Computational Geometry. pp 326-333.
DEHNE, F., AND NOLTEMEIER, H. 1985. A computa-
tional geometry approach to clustering prob- EDAHIRO, M., KOKUBO, I., AND ASANO, T. 1984. A
lems. In Proceedings of the Ist Annual ACM new point-location algorithm and its practical
Symposium on Computational Geometry. pp. efficiency: Comparison with existing algo-
245-250. rithms. ACM Trans. Graph. 3, 86-109.
DELAUNAY, B. N. 1932. Neue Darstellung der ge- EDELSBRUNNER, H 1986 Edge-skeletons in ar-
ometrischen Kristallographie. Z. Kristallo - rangements with applications. AlgorLthmica 1,
graph. 84, 109-149. 93-109,
DELAUNAY, B. N. 1934. Sur la sphere vide, Bull. EDELSBRUNNER, H. 1987. Algorithms in Combina-
Acad. Science USSR VII: Class. Sci. Math., torial Geometry. Springer, Berlin-Heidelberg.
793-800. EDELSBRUNNER, H. 1989. An acyclicity theorem
DELAUNAY, B. N. 1963. Theorie der regularen for cell complexes in d dimensions. In Proceed-
Dirichletschen Zerlegungen des n-dimensiona- ings of the 5th Annual ACM Symposium on
len euklidischen Raumes. SchriftenreLhe Int. Computational Geometry. pp. 145-151.
Math. Dtsch. Akad. Wiss. Berlin 13, 27-31. EDELSBRUNNER, H., AND MAURER, H. A. 1985.
DELAUNAY, B. N., DOLBILIN, N. P., AND STOGRIN, Finding extreme points in three dimensions
M. I. 1978. Combinatorial and metric theory and solving the post-office problem in the plane.
of planigons. Trudy Mat. Inst. Steklov 148, Inf Process. Lett. 21, 39-47.
109-140. EDELSBRUNNER, H., AND SEIDEL, R. 1986. Voronoi
DEWDNEY, A. K. 1977. Complexity of nearest diagrams and arrangements. Discrete Comput.
neighbour searching in three and higher di- Geom. 1, 25-44.
mensions. Rep. 28, Univ. of Western Ontario, EDELSBRUNNER, H., AND WELZL, E. 1985. On the
London, Ontario. number of line separations of a finite set in the
DEWDNEY, A. K., AND VRANKXI, J. K. 1977. A plane. J. Combin. Theory Ser. A, 15-29.
convex partition of R 3 with applications to EDELSBRUNNER, H., GUIBAS, L. J., AND SHARIR, M.
Crum’s problem and Knuth’s post-oftlce prob- 1989. The upper envelope of piecewise linear
lem. Util. Math. 12, 193-199. functions: algorithms and applications.
DILLENCOURT, M. B. 1987a. Traveling salesman Discrete Comput. Geom. 4, 311-336.
cycles are not always subgraphs of Delaunay EDELSBRUNNER, H., GUIBAS, L. J., STOLFI, J. 1986.
triangulations. Infi Process. Lett. 24, 339-342. Optimal point location in a monotone subdivi-
DILLENCOURT, M. B. 1987b. A non-Hamiltonian, sion. SIAM J. Comput. 15, 317-340.
nondegenerate Delaunay triangulation. Infi EDELSBRUNNER, H., KIRKPATRICK, D. G., AND
Process. Lett. 25, 149-151. SEIDEL, R. 1983. On the shape of a set of

points in the plane. IEEE Trans. Znf Theory GAUSS, C. F. 1840. Recursion der ‘Untersuchun-
IT-29, 551-559. gen iiber die Eigenschaften der positiven
EDELSBRUNNER, H., O’ROURKE, J., AND SEIDEL, R. ternaren quadratischen Formen von Ludwig
1986. Constructing arrangements oflines and August Seeber. J. Reine Angew. Math. 20,
hyperplanes with applications. SZAMJ. Com- 312-320.
put. 15,341-363. GILBERT, E. N. 1962. Random subdivisions of
EDELSBRUNNER, H., ROTE, G., AND WELZL, E. 1987. space into crystals. Ann. Math. Stat. 33,
Testing the necklace condition for shortest 958-972.
tours andoptimal factors inthe plane. Springer GOODRICH, M. T., O’DUNLAING, C., AND YAP, C. K.
LNCS 267, 364-375. 1989. Constructing the Voronoi diagram of a
EHRLICH, P. E., AND IMHOF, H.C. 1979. Dirichlet set of line segments in parallel. Springer LNCS
regions in manifolds without conjugate points. 382, 12-23.
Comment. Math. Helvet. 54,642-658. GOWDA, I. G., KIRKPATRICK, D. G., LEE, D. T., AND
EISELT, H. A., AND PEDERZOLI, G. 1986. Voronoi NAAMAD, A. 1983. Dynamic Voronoi dia-
diagrams andtheir use: A survey. Part I: The- grams. IEEE Trans. Infi Theory IT-29,
OrY; Part 11: Applications. In S. Goyal, Ed. 724-731.
Proceedings pp. 98-115.
of the ASAC 7, 2. GREEN, P. J., AND SIBSON, R. 1977. Computing
ENGEL,P. 1981. ~er Wirkungsbereichsteilungen Dirichlet tessellations in the plane. Comput.
mit kubischer Symmetrie. Z. Kristallograph. J. 21, 168-173.
154, 199-215. GRUBER, P. M., AND LEKKERKERKER, C. G. 1988.
EPPSTEIN, D. 1990. Finding the k smallest span- Geometry of Numbers. North-Holland Publish-
ning trees. Springer LNCS447, 38-47. ing Co., Amsterdam.
FAIRFIELD, J. 1979. Contoured shape generation: GRUBER, P. M., AND RYSKOV, S. S. 1987. Facet-to-
Forms that people seeindot patterns. In Proc- facet implies face-to-face. Manuscript.
eedings of the IEEE Conference on Systems Man
GRUNBAUM, B. 1967. Convex Polytopes. Inter-
Cybernetics. pp. 60-64. science, New York.
FAIRFIELD,J. 1983. Segmenting dot patterns by
GRtiNBAUM, B., AND SHEPHARD, G. C. 1980. Tilings
Voronoi diagram concavity. IEEE Trans. Patt.
with congruent tiles. Bull. Am. Math. Sot. 3,
Anal. Mach. Znt. PAMZ-5, 104-110.
951-973.
FEDEROFF, E. S. 1885. Elemente der Lehre von
GRUNBAUM, B., AND SHEPHARD, G. C. 1987. Tilings
den Fignren. Verb. Russ Min. Ges. St. Peters-
and Patterns. Freeman and Co., New York.
burg 21, 1-279.
GUIBAS, L. J., AND STOLFI, J. 1985. Primitives for
FIELD, D. A. 1986. Implementing Watson’s algo-
the manipulation of general subdivisions and
rithm in three dimensions. In Proceedings of
the computation of Voronoi diagrams. ACM
the 2nd Annual ACM Symposium on Computa-
Trans. Graph. 4, 74-123.
tional Geometry. pp. 246-259.
FINNEY, J. L. 1979. Procedure for the construction GUIBAS, L. J., KNUTH, D. E., AND SHARIR, M. 1990.
of Voronoi polyhedra. Note. J. Comput. Phys. Randomized incremental construction of
Delaunay and Voronoi diagrams. Springer
32, 137-143.
LNCS 443, 414-431.
FORTUNE, S. 1985. A fast algorithm for polygon
containment by translation. Springer LNCS GUIBAS, L. J., MITCHELL, J. S. B., AND Roos, T.
194, 189-198. 1991. Voronoi diagrams of moving points in
the plane. Springer LNCS. In press.
FORTUNE, S. 1987. A sweepline algorithm for
Voronoi diagrams. Algorithmic 2, 153-174. HARTIGAN, J. A. 1975. Clustering Algorithms.
John Wiley, New York.
FORTUNE, S. 1988. Sorting helps for Voronoi dia-
grams. Manuscript, ATT Bell Lab., Murray HEUSINGER, H., AND NOLTEMEIER, H. 1989. On
Hill, N.J. separable clusterings. J. Algorithms 10,
212-227.
FRANK, F. C., AND KASPER, J. S. 1958. Complex
alloy structures regarded as sphere packings. HINRICHS, K., NIEVERGELT, J., SCHORN, P. 1988a.
Acts Crystallogr. 11, 184-190. Plane-sweep solves the closest pair problem el-
egantly. In~ Process. Lett. 26, 255-261.
GABRIEL, K. R., AND SOKAL, R. R. 1969. A new
statistical approach to geographic variation HINRICHS, K., NIEVERGELT, J., SCHORN, P. 1988b.
analysis. Syst. Zoology 18, 259-278. A sweep algorithm for the all-nearest-
GAMBINI, R. 1966. A computer program for calcu- neighbors problem. Springer LNCS 333, 43–54.
lating lines of equilibrium between multiple HORTON, R. E. 1917. Rational study of rainfall
centers of attraction. Manuscript, Center of Re- data makes possible better estimates of water
gional Studies, Univ. of Kansas, Lawrence, yield. Eng. News-Record, 211-213.
Kan. HOWE, S. E. 1978. Estimating regions and clus-
GAREY, M. R., GRAHAM, R. L., AND JOHNSON, D. S. tering spatial data Analysis and implementa-
1976. Some NP-complete geometric prob- tion of methods using the Voronoi diagram.
lems. In Proceedings of the 8th Annual ACM Ph.D. dissertation, Brown Univ., Providence,
Symposium on STOC. pp. 10-22. R.I.

HUTTENLOCHER, D. P., KEDEM, K., AND SHARIR, M. LANTUEJOUL, C., AND MAISONNEUVE, F. 1984.
1991. Theupper envelope of Voronoi surfaces Geodesic methods in quantitative image analy-
and its applications. In Proceedings of the 7th sis. Pattern Recogn. 17, 177-187.
Annual ACM Symposium on Computational LAVES, F. 1930. Ebeneneinteilung in Wirkungs -
Geometry, 194-203. bereiche. Z. Kristallograph. 76, 277-284.
HWANG, F. K. 1979. An O(nlogn) algorithm for LAWSON, C. L. 1972. Generation of a triangular
rectilinear minimal spanning trees. ~. ACM grid with applications to contour plotting. Tech.
26, 177-182. Memorandum 299, California Inst. Tech. Jet
IMAI, H., IRI, M., AND MUROTO, K. 1985. Voronoi Propulsion Lab.
diagram in the Laguerre geometry and its ap- LAWSON, C. L. 1977. Software for Cl surface
plications. SIAM J. Comput. 14, 93-105. interpolation. In Mathematical Software III,
IMAI, K., SUMmO, S., IMAI, H. 1989. Minimax geo- J. Rice Ed., Academic Press, New York.
metric fitting of two corresponding sets of LEE, D. T. 1980. Two-dimensional Voronoi dia-
points. In Proceedings of the 5th Annual ACM grams in the Lp-metric. J. ACM 27, 604-618.
Symposium on Computational Geometry.
LEE, D. T. 1982a. On k-nearest neighbor Voronoi
pp. 266-275.
diagrams in the plane. IEEE Trans. Comput.
JOHNSON, W. A., AND MEHL, R. F. 1939. Reaction (X31, 478-487.
kinetics in processes of nucleation and growth. LEE, D. T. 1982b. Medial axis transform of a pla-
Trans. Am. Instit. Mining Metall. A. I. M.M.E. nar shape. IEEE Trans. Putt. Anal. Mach.
135, 416-458. Intell. PAM14, 363-369.
KANTABUTEA, V. 1983. Traveling salesman cycles LEE, D. T., AND DRYSDALE, R. L. 1981. General-
are not always subgraphs of Voronoi duals. ization of Voronoi diagrams in the plane. SIAM
Znfi Process. Lett. 16, 11-12. J. Comput. 10, 73-87.
KEIL, J. M., AND GUTWIN, C. A. 1989. The Delau- LEE, D. T., AND LIN, A. K. 1986. Generalized
nay triangulation closely approximates the Delaunay triangulation for planar graphs. Dis-
complete Euclidean graph. Springer LNCS 382, crete Comput. Geom. 1, 201-217.
47-56.
LEE, D. T., AND PREPARATA, F. P. 1984. Computa-
KLANG, T. 1966. Random fragmentation in two tional geometry—A eurvey. IEEE Trans. Com -
and three dimensions. Z. Astrophysik 64, put. C33, 12, 1072-1101.
433-439.
LEE, D. T., AND SCHACHTER, B. J. 1980. Two algo-
KDtKPATRICK, D. G. 1979. Efi-lcient computation rithms for constructing a Delaunay triangula-
of continuous skeletons. In Proceedings of the tion. Int. J. Comput. Inf Sci. 9, 219-242.
20th Annual IEEE Symposium on FOCS.
LEE, D. T., AND WONG, C. K. 1980. Voronoi
pp. 18-27.
diagrams in the Ll( La) metrics with 2-
KmKPATRICK, D. G. 1980. A note on Delaunay dimensional storage applications. SIAM J.
and optimal triangulations. Inf Process. Lett. Comput. 9, 200-211.
10, 127-128.
LEE, D. T., AND YANG, C. C. 1979. Location of
KIRKPATRICK, D. G. 1983. Optimal search in pla- multiple points in a planar subdivision. Znf
nar subdivisions. SIAM J. Comput. 12, 28-35. Process. Lett. 9, 190-193.
KLEE, V. 1980. On the complexity of d-dimen- LEVEN, D., AND SHARIR, M. 1986. Intersection and
sional Voronoi diagrams. Archiu Math. 34, proximity problems and Voronoi diagrams.
75-80. Adu. Robotics 1, 187-228.
KLEIN, R. 1989. Concrete and Abstract Voronoi LEVIN, D., AND SHARIR, M. 1987. Planning a
D~agrams. Springer LNCS 400. purely translational motion of a convex robot
KLEIN, R., AND WOOD, D. 1988. Voronoi diagrams in two-dimensional space using Voronoi dia-
based on general metrics in the plane. Springer grams. Discrete Comput. Geom. 2, 9-31.
LNCS 294, pp.281-291, LINGAS, A. 1986a. The greedy and Delaunay tri-
KNUTH, D. 1973. The Art of Computer Program- angulations are not bad in the average case.
ming III: Sorting and Searching. Addison- Infi Process. Lett. 22, 25-31.
Wesley, Reading, Mass. LINGAS, A. 1986b. A fast algorithm for the gener-
KOCH, E. 1973. Wirkungsbereichspolyeder und alized Delaunay triangulation. Manuscript,
Wirkungsbereichsteilungen zu kubischen Git- Univ. of Linkoping, Sweden.
terkomplexen mit weniger als drei Freiheits- LINGAS, A. 1989. Voronoi diagrams with barriers
graden. Z. Kristallograph. 138, 196-215. and the shortest diagonal problem. lnfi Pro-
KOPEC, R. J. 1963. An alternative method for the cess. Lett. 32, 191-198.
construction of Thiessen polygons. Professional LINHART, J. 1981. Die Beleuchtung von Kugeln.
Geographer 15, 24-26. Geom. Dedicata 10, 145-154.
KRUSKAL, J. B. 1956. On the shortest spanning LITTLE, D, V. 1974. A third note on recent re-
subtree of a graph and the traveling salesman search in geometric probability. Adu. Appl.
problem. Problem. Am. Math. Sot. 7, 48-50. Probab. 6, 103-130.

LOEB, A. L. 1970. A systematic survey of cubic in O(n log n) time. In Proceedings of the 4th
crystal structures. J. Solid State Chem. 1, Annual ACM Symposium on Computational
237-267. Geometry. pp. 143-152.
MANACHER, G. K., AND ZOBRIST, A. L. 1979. Nei- MUDER,D. J. 1988a. How big is an n-sided
ther the greedy nor the Delaunay triangula- Voronoi polygon? Manuscript, MITRE Corp.,
tion of a planar point set approximates the Bedford,Mass.
optimal triangulation. lnfi Process. Lett. 9, MUDER,D. J. 1988b. Putting the best face on a
31-34. Voronoi polyhedron. Proc. London Math. Sot.
MATULA, D. W., AND SOKAL, R. R. 1980. Proper- 56, 329-348.
ties of Gabriel graphs relevant to geographic MULMULEY, K. 1991. On levels in arrangements
variation research and the clustering of points and Voronoi diagrams. Discrete Comput. Geom.
in the plane. Geograph. Anal. 12, 205-221. In press.
MATZKE, E. B., AND NESTLER, J. 1946. Volume- MURTAGH, F. 1983. A survey of recent advances
ehape relationships in variant foams. Am. J. in hierarchical clustering algorithms. Com-
Botcmics 33, 130-144. puter J. 26, 354-359.
MAUS, A. 1884. Delaunay trianwlation and the NEWMAN, C. M., RINOTT, Y., AND TVERSKY, A. 1983.
convex hull of n point; in expected linear time. Nearest neighbor and Voronoi regions in cer-
BZT24, 151-163. tain point processes. Adv. Appl. Probab. 15,
MAXWELL, J. C. 1864. On reciprocal diagrams and 726-751.
diagrams of forces. Phil. Mug. 4, 27, 250-261. NIGGLI, R. 1927. Die topologische Strukturanal.
MCLAIN, D. H. 1976. Two dimensional interpola- yse. Z. Kristallograph. 65, 391-415.
tion from random data. Comput. J. 19, NOWACKI, W. 1933. Der Begriff ‘Voronoischer
178-181. Bereich’. Z. Kristallograph. 85, 331-332.
MEGIDDO, N. 1983. Linear time algorithms for lin- NOWACKI, W. 1976. Uber allgemeine Eigen-
ear programming in R 3 and related problems. schaften von Wirkungsbereichen. Z. Kristallo -
SIAM J. Comput. 12, 759-776. graph. 143, 360-368.
MEHLHORN, K., O’DUNLAING, C., AND MEISER, S. O’DUNLAING, C., AND YAP, C. K. 1985. A “retrac-
1990. On the construction of abstract Voronoi tion” method for planning the motion of a disc.
diagrams. Springer LNCS 415, 227-239. J. Algorithms 6, 104-111.
MEIJERING, J. L. 1953. Interface area, edge length, O’DUNLAING, C., SHARIR, M., AND YAP, C. K. 1986.
and number of vertices in crystal aggregates Generalized Voronoi diagrams for moving a
with random nucleation. Philips Res. Rept. 8, ladder: I. Topological analysis. Comm. Pure
270-290. Appl. Math. 39, 423-483.
MILES, R. E. 1970. On the homogeneous planar O’DUNLAING, C., SHARIR, M., AND YAP, C. K. 1987.
Poisson process. Math. Biosci. 6.85-127. Generalized Voronoi diagrams for moving a
MITCHELL, J. S. B., MOUNT, D. M., AND ladde~ II. Eftlcient construction of the dia-
PAPADIMITRIOU, C. H. 1987. The discrete gram. Algorithmic 2, 27-59.
geodesic problem. SIAM J. Comput. 16, OHYA, T., IRI, M., AND MUROTA, K. 1984a. A fast
647-668. Voronoi diagram algorithm with quaternary
MOLLISON, D. 1977. Spatial contact models for tree bucketing. Infi Process. Lett. 18, 227-231.
ecological and epidemic spread. J. Roy. Stat. OHYA, T., I.RI, M., AND MUROTA, K. 1984b. Im-
SOC. B 39, 283-326. provements of the incremental methods for the
MONTANARI, U. 1968. A method for obtaining Voronoi diagram with computational compari-
skeletons using a quasi. Euclidean distance. son of various algorithms. J. Operations Res.
J. ACM 15, 600-624. Sot. Japan 27, 306-337.
MORAN, P. A. P. 1966. A note on recent research OKABE, A., YOSHIKAWA, T., FUJII, A., AND OIKAWA,
in geometric probability. J. Appl. Probab. 3, K. 1988. The statistical analysis of a distribu.
453-463. tion of active points in relation to surface-like
MORAN, P. A. P. 1969. A second note on recent elements. Environ. Plan. A 20, 609-620.
research in geometrical probability. Adu. Appt. OVERMARS, M. 1981. Dynamization of order de-
Probab. 1, 73-89. composable set problems. J. Algorithms 2,
MOUNT, D. M. 1985. Voronoi diagrams on the 245-260.
surface of a polyhedron. Rep. 1496, Univ. of PACH, J., AND SHARIR, M. 1989. The upper enve-
Maryland. lope of piecewise linear functions and the
boundary of a region enclosed by convex plates.
MOUNT, D. M. 1986. Storing the subdivision of a
Discrete Comput. Geom. 4, 291-310.
polyhedral surface. In Proceedings of the 2nd
Annual ACM Symposium on Computational PAPADIMITRIOU, C. H. 1977. The Euclidean travel-
Geometry. pp. 150-158. ing salesman problem is NP-complete. ‘lheor.
MOUNT, D., AND SAALFELD, A. 1988. Globally- Comp. Sci. 4, 237-244.
equiangular triangulation of co-circular points PASCHINGER, I. 1982. Konvexe Polytope und

Dirichletsche Zellenkomplexe. Ph.D. Disserta- grams in higher dimensions. In Proceedings of

tion, Inst. f. Math., Univ. Salzburg, Austria. the 20th Annual Allerton Conference on CCC.
PHILBRICK, O. 1968. Shape recognition with the pp. 94-95.
medial axie transform. In Pzctorial Pattern SEIDEL, R. 1985 A method for proving lower
Recognition, G. C. Cheng, R. S. Ledley, D. K bounds for certain geometric problems. In
Pollock, and A. Rosenfeld, Eds. Washington, Computational Geometry, G. T. Toussaint Ed.
D.C. pp. 319-334.
PREPARATA, F. P. 1977. Steps into computational SEIDEL, R. 1986. Constructing higher-dimensional
geometry. Rep. R-760, Coordinated Science convex hulls at logarithmic cost per face. In
Lab., Univ. of Illinois, Urbana, Ill., pp. 23-24. Proceedings of the 18th Annual ACM Sympo-
PREPARATA, F. P., AND HONG, S. J. 1977. Convex sium on STOC. pp. 404-413.
hulls of finite sets of points in two and three SEIDEL, R. 1987. On the number of faces in
dimensions. Commun. ACM 20, 87-93. higher-dimensional Voronoi diagrams. In Pro-
PREPARATA, F. P., AND SHAMOS, M. I. 1985. Com- ceedings of the 3rd Annual Symposium on
putational Geometry: An Introduction. Computational Geometry. pp. 181-185.
Springer, New York. SEIDEL, R. 1988. Constrained Delaunay triangula-
PRIM, R. C. 1957. Shortest connection networks tions and Voronoi diagrams with obstacles. In
and some generalizations Bell Syst. Tech. J. Rep. 260, IIG-TU Graz, Austria, pp. 178-191.
36, 1389-1401. SHAMOS, M. I. 1975. Geometric complexity. In
RAJAN, V. T, 1991. 0 timality of the Delaunay Proceedings of the 7th Annual ACM Sympo-
1 .sLum on STOC. pp. 224-233.
triangulation in R In Proceedings of the 7th
Annual ACM Symposmm on Computat~onal SHAMOS, M. I. 1978. Computational geometry.
Geometry, pp. 357-372. Ph.D. dissertation, Yale Univ., New Haven,
RHYNSBURGER, D. 1973. Analytical delineation of Corm.
Thiessen polygons. Geograph. Anal. 5, SHAMOS, M. I., AND HOEY, D. 1975. Closest-point
133-144. problems. In Proceedings of the 16th Annual
RIPPA, S. 1990. Minimal roughness property of IEEE Sympostum on FOCS pp. 151-162.
the Delaunay triangulation. Comput. Aided SHARIR, M. 1985. Intersection and closest-pair
Geom. Des~gn 7, 489-49’7. problems for a set of planar discs. SIAM J.
ROGERS, C. A. 1964. Packing and Covering. Cam- Comput. 14, 448-468.
bridge University Press, London, New York. SIBSON, R. 1977. Locally equiangular triangula-
ROHNERT, H. 1988. Moving discs between poly- tions. Comput. J. 21, 243-245.
gons. Springer LNCS 317, 502-515. SIBSON, R. 1979. The Dirichlet tessellation as an
ROSENBERGER, H. 1988. Order-k Voronoi dia- aid in data analysis. Scandinavian J. Stat. 7,
grams of sites with additive weights in the 14-20.
plane. Rep. UIUCDCS-R-88-1431, Dept. Com- SIBSON, R. 1980. A vector identity for the Dn-ich-
put. Sci., Univ. Illinois, Urbana, Ill. let tessellation. Math. Proc. Carob. Phil. Sot.
ROSENKRANTZ, D. J., STEARNS, R. E., AND LEWIS, P. 87, 151-155.
M. 1977. An analysis of several heuristics for SmRONY, S., AND SHARIR, M. 1986. A new efficient
the traveling salesman problem. SIAM J. motion-planning algorithm for a rod in polygo-
Comput. 6, 563-581. nal space. In Proceedings of the 2nd Annual
ROWAT, P. F. 1979. Representing spatial experi- ACM Symposium on Computational Geometry.
ence and solving spatial problems in a simu- pp. 178-186.
lated robot environment. Ph.D. dissertation, SMITH, C. S, 1954. The shape of things. Sci. Am.
Univ. of British Columbia, Canada. 58-64.
SAKAMOTO, M., AND TAKAGI, M. 1988. Patterns of SNYDER, D. E. 1962. Urban places in Uruguay
weighted Voronoi tessellations. Scl. Form 3, and the concept of a hierarchy. Festschrift:
103-111. C. F. Jones. Northwestern University Studies
SCHONFLIES, A. 1891. Kristallsysteme und in Geography 6, 29-46.
Kirstallstruktur. Teubner, Leipzig. STIFTER, S. 1989. An axiomatic approach to
SCHWARTZ, J., AND YAP, C. K. 1986. Advances in Voronoi diagrams in 3D. Manuscript, RISC,
Robotics. Lawrence Erlbaum Associates, Hill- Univ. of Linz, Austria.
side, N.J. SUGIHARA, K., AND IRI, M. 1988. Geometric algo-
SCHWARZKOPF, O. 1989. Parallel computation of rithms in finite-precision arithmetic. Research
discrete Voronoi diagrams. Springer LNCS Memorandum RMI 88-10, Faculty of Engineer-
349, 193-204. ing, Univ. of Tokyo, Japan.
SEIDEL, R. 1981. A convex hull algorithm optimal SUPOWIT, K. J. 1983. The relative neighborhood
for point sets in even dimensions. Rep. 81-14, graph, with an application to minimum span-
Univ. of Britmh Columbia, Vancouver, Canada. ning trees. J. ACM 30, 428-448.
SEIDEL, R. 1982 The complexity of Voronoi dia- SUZUKI, A., AND IRI, M. 1986. Approximation of a

tessellation of the plane by a Voronoi diagram. quadratiques. J. Reine Angew. Math. 134,
J. Operations Res. Sot. Japan 29, 69-96. 198-287.
TANEMURA, M., OGAWA, T., AND OGITA, N. 1983. A WANG, C. A., AND SCHUBERT, L. 1987. An optimal
new algorithm for three-dimensional Voronoi algorithm for constructing the Delaunay trian-
tessellation. J. Comput, l%ys. 51, 191-207, gulation of a set of line segments. In Proceed-
TARSKI, A. 1951. A decision method for elemen- ings of the 3rd Annual ACM Symposium on
tary algebra and geometry. Univ. of California Computational Geometry. pp. 223-232.
Press, Berkeley, Calif. WATSON, D. F. 1981. Computing the n-dimen-
THIESSEN, A. H. 1911. Precipitation average for sional Delaunay tessellation with application
large area. Monthly Weather Rev. 39, to Voronoi polytopes. Comput. J. 24, 167-172.
1082-1084. WEAIRE, D., AND RIVIER, N. 1984. Soap, cells, and
TOKUYAMA, T. 1988. Deformation of merged statistics—random patterns in two dimensions.
Voronoi diagrams with translation. Rep. TR- Contemp. Phys. 25, 59-99.
88-0049, IBM Tokyo Research Lab. WHITELEY, W. 1979. Realizability of polyhedra.
TOUSSAINT, G. T. 1980. The relative neighborhood Structural Topol. 1, 46-58.
graph of a finite planar set. Pattern Recogn. WIDMEIER, P., Wu, Y. F., AND WONG, C. K. 1987.
12, 261-268. On some distance problems in fixed orienta-
TOUSSAINT, G. T., AND BHATTACHARYA, B. K. 1981. tions. SIAM J. Comput. 16, 728-746.
On geometric algorithms that use the furthest- WIGNER, E., AND SEITZ, F. 1933. On the constitu-
point Voronoi diagram. Rep. 81-3, McGill tion of metallic sodium. Phys. Rev. 43,
Univ., Montreal, Quebec, Canada. 804-810.
TUOMINEN, O. 1949. Das Einflu13gebeit der Stadt WILLIAMS, R. E. 1968. Space-filling polyhedron:
Turku ins System der Einflut3gebiete Its relation to aggregates of soap bubbles, plant
SW–Finnlands. Fennia 71-5, 1-138. cells, and metal crystalites. Science 161,
URQUHART, R. 1980. A note on the computation of 276-277.
subgraphs of the Delaunay triangulation. Rep., YAO, A. C. 1975. An O(E log log V) algorithm for
Univ. of Glasgow, U.K. finding minimum spanning trees. Znfi Process.
VAIDYA, P. M. 1988a. Minimum spanning trees in Lett. 4, 21-23.
k-dimensional space. SIAM J. Comput. 17, YAO, A. C. 1982. On constructing minimum span-
572-582. ning trees in k-dimensional spaces and related
VAIDYA, P. M. 1988b. Geometry helps in match- problems. SIAM J. Comput. 11, 721-736.
ing. In Proceedings of the 20th Annual ACM YAP, C. K. 1987. An 0( n log n) algorithm for the
Symposium on STOC. pp. 422-425. Voronoi diagram of a set of simple curve seg-
VORONOI, M. G. 1908. Nouvelles applications des ments. Discrete Comput. Geom. 2, 365-393.
parametres continus a la theorie des formes
Received February 1989, final rewslon accepted January 1991.
ACM Computing Surveys, Vol. 2.3, No. 3, September 1991

Voronoi Diagrams - A Survey of A Fundamental Geometric Data Structure

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Voronoi Diagrams - A Survey of A Fundamental Geometric Data Structure

Uploaded by

Copyright:

Available Formats

Voronoi Diagrams — A Survey of a Fundamental Geometric

Categories and Subject Descriptors: F.2.2 [Analysis of Algorithms and Problem

General Terms: Algorithms, Theory

INTRODUCTION [19841 and to the textbooks by Preparata

ACM Computing Surveys, Vol. 23, No. 3, September 1991

CONTENTS developed for the computer construction

We begin with a description of elemen-

ACM Computmg Surveys, Vol. 23, No. 3, September 1991

should be observed that each vertex is

Consider some file of n two-attribute

ACM Computmg Surveys, Vol. 23, No 3, September 1991

income of a person. Suppose we are given

ACM Computing Surveys, Vol 23, No 3, September 1991

in the L1-metric (i. e., Manhattan dis-

cessing batched requests is the following:

ACM Computing Surveys, Vol. 23, No. 3, September 1991

The history of’ Voronoi diagrams can be

(1) Their use in modeling natural phe-

Accordingly, Voronoi diagrams are Figure 5. Cubic crystal structure.

ACM Computing Surveys, Vol. 23, No. 3, September 1991

ACM Computing Surveys, Vol. 23, No. 3, September 1991

ACM Computing Surveys, Vol. 23, No. 3, September 1991

Figure 7. Spider web

Figure 9. Contour and its medial axis [Philbrick

exploited by Dirichlet [18501 to establish,

ACM Computing Surveys, Vol. 23, No. 3, September 1991

these terms are the most popular ones in

1.2.1 Regularly Placed Sites

Mathematical interest in Voronoi dia-

ACM Computing Surveys, Vol. 23, No. 3, September 1991

and covering problems for congruent

2 In fact, a tiling if facet to facet if and only if it is

ACM Computing Surveys, Vol. 23, No 3, September 1991

Figure 12. Apollomus model [Aurenhammer and Edelsbrunner 1984].

62( x, p) – w. Intuitively speaking, w ex-

the geometry of numbers [Gruber and [Shames19781.

ACM Computing Surveys, Vol 23, No 3, September 1991

ACM Computing Surveys, Vol. 23, No. 3, September 1991

angulations without “extreme” an-

ACM Computmg Surveys, Vol. 23, No 3, September 1991

ACM Computing Surveys, Vol. 23, No. 3, September 1991

(generalized diagrams), Besag [19741, intuitive construction rules were pro-

ACM Computing Surveys, Vol 23, No 3, September 1991

diagram in R d that grows exponentially

the number of edges deleted. This num- 1.3.2 Speeding Up Insertion

ACM Computmg Surveys, Vol. 23, No. 3, September 1991

Voronoi diagrams by Klein [1989] with-

an even more practical version of ran- -.. .:.’---,

the construction of Voronoi diagrams. For /

ACM Computing Surveys, Vol 23, No 3, September 1991

ACM Computing Surveys, Vol. 23, No. 3, September 1991

space (for any c > O) that is nearly

Figure 18. Maintaining the intersection with a

Another powerful technique in computa-

ACM Computing Surveys, Vol. 23, No. 3, September 1991

triangulation are maintained to facili-

ACM Computing Surveys, Vol 23, No. 3, September 1991

computing Voronoi diagrams, such as di- Delaunay triangulation (Figure 21).

ACM Computing Surveys, Vol. 23, No. 3, September 1991

Figure 20. Voronoi tree [Gowda et al. 19831.

ACM Computing Surveys, Vol. 23, No 3, September 1991

Figure 21. Delaunay tree.

I I In her pioneering work, Chow [1980]