Computational Geometry: Theory and Applications: Rodrigo I. Silveira, Marc Van Kreveld

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Computational Geometry 42 (2009) 803813

Contents lists available at ScienceDirect

Computational Geometry: Theory and Applications


www.elsevier.com/locate/comgeo

Optimal higher order Delaunay triangulations of polygons


Rodrigo I. Silveira , Marc van Kreveld
Department of Information and Computing Sciences, Utrecht University, the Netherlands

a r t i c l e

i n f o

Article history:
Received 18 May 2007
Received in revised form 18 December 2007
Accepted 20 February 2008
Available online 25 February 2009
Communicated by O. Aichholzer and
F. Aurenhammer
Keywords:
Optimal triangulation
Higher order Delaunay triangulations
Optimization

a b s t r a c t
This paper presents an algorithm to triangulate polygons optimally using order-k Delaunay
triangulations, for a number of quality measures. The algorithm uses properties of higher
order Delaunay triangulations to improve the O (n3 ) running time required for normal
triangulations to O (k2 n log k + kn log n) expected time, where n is the number of vertices
of the polygon. An extension to polygons with points inside is also presented, allowing
to compute an optimal triangulation of a polygon with h  1 components inside in
O (kn log n) + O (k)h+2 n expected time. Furthermore, through experimental results we show
that, in practice, it can be used to triangulate point sets optimally for small values of k. This
represents the rst practical result on optimization of higher order Delaunay triangulations
for k > 1.
2009 Elsevier B.V. All rights reserved.

1. Introduction
One of the best studied topics in computational geometry is the triangulation. When the input is a point set P , it is
dened as a subdivision of the plane whose bounded faces are triangles and whose vertices are the points of P . When the
input is a polygon, the goal is to decompose it into triangles by drawing diagonals.
Triangulations have applications in a large number of elds, including computer graphics, multivariate analysis, mesh
generation, and terrain modeling. Since for a given point set or polygon, many triangulations exist, it is possible to try to
nd one that is the best according to some criterion that measures some property of the triangulation.
The properties of interest are application-dependent, but are generally either local properties of the triangles (like area,
height or minimum angle) or global properties of the triangulation (such as total edge length). For example, in automatic
mesh generation for nite element methods, criteria like minimizing the minimum/maximum angle and height of the
triangles are related to the error of the nite element approximation [3,21]. In the context of terrain modeling, terrains
are many times represented by triangulated irregular networks, which are triangulations where each point has an elevation.
When terrain models need to be realistic, for example in visualization or terrain analysis for hydrology, criteria like avoiding
ill-shaped triangles and having few local minima are particularly relevant [6,18].
For a given set of points P , a well-known triangulation is the Delaunay triangulation. It is dened as a triangulation
where the circumcircle of the three vertices of any triangle does not contain any other point of P . It is unique when no
four points are cocircular, and can be computed in O (n log n) time for n points. The Delaunay triangulation has a large
number of known properties, and it optimizes several measures, like max min angle or min max smallest enclosing circle,
among others. This is the reason why its triangles are said to be well-shaped. However, for many applications, the Delaunay
triangulation is not exible enough. For example, when used for terrain modeling, the triangulation does not take the third
dimension into account, which may result in artifacts like interrupted valley lines or articial local minima.

This research has been partially funded by the Netherlands Organisation for Scientic Research (NWO) under the project GOGO.
Corresponding author.
E-mail addresses: [email protected] (R.I. Silveira), [email protected] (M. van Kreveld).

0925-7721/$ see front matter 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.comgeo.2008.02.006

804

R.I. Silveira, M. van Kreveld / Computational Geometry 42 (2009) 803813

Fig. 1. A Delaunay triangulation (k = 0) (left), and an order-2 triangulation (right). Light gray triangles are rst order, the medium gray ones are second
order.

Fig. 2. Three different triangulations of a polygon: the constrained Delaunay triangulation (left), the minimum weight triangulation (center), and the minimum weight triangulation constrained to order-2 Delaunay triangulations (right). The third one combines nicely shaped triangles with the minimization of
the weight.

To overcome this limitation, Gudmundsson et al. [11] dene higher order Delaunay triangulations, a class of well-shaped
triangulations where a few points are allowed inside the circumcircles of the triangles. A triangulation is order-k Delaunay
if the circumcircle of the three vertices of any triangle contains at most k other points (see Fig. 1). For k = 0, each nondegenerate point set has only one higher order Delaunay triangulation, equal to the Delaunay one. As the parameter k is
increased, more points inside the circumcircles imply a reduction in the shape quality of the triangles, but also an increase
in the number of triangulations that are considered, and hence greater exibility to optimize some other criterion, while
limiting the badness of the shape of the triangles. The concept of higher order Delaunay triangulation has been successfully
applied to several areas, including terrain modeling [7], minimum interference networks [1] and multivariate splines [20].
In this paper we focus mainly on triangulations of polygons. Optimal polygon triangulation has been a subject of study
for a long time, both because it has applications of its own, like in nite element methods, and also because it gives insight
into the generally harder problem of optimal point set triangulation. When the goal is to optimize only one criterion, optimal
polygon triangulations can be computed in polynomial time for many measures. The constrained Delaunay triangulation [5],
which generalizes the standard denition in order to force certain edges into the triangulation, can be used to triangulate
polygons, with triangle shape properties similar to the ones of the Delaunay triangulation. Many other measures can be
optimized using a dynamic programming algorithm attributed to Klincsek [16], and also, independently, proposed by Gilbert
[10]. This approach allows to nd in O (n3 ) time an optimal triangulation of a simple polygon for any decomposable measure.
Intuitively, a measure is decomposable if the measure of the whole triangulation can be computed eciently from the
measures of two pieces, together with the information on how the pieces are glued together. See [3] for a formal denition.
Decomposable measures include the following: min or max angle, min or max circumcircle, min or max length of an edge,
min or max area of triangle, and the sum of the edge lengths. The algorithm by Klincsek can be extended to other measures
that are not decomposable, like maximum vertex degree. For convex polygons, the min or max area of triangle measures
can be optimized even faster, in O (n2 log n) time [15].
Triangulating point sets optimally is in general more dicult than triangulating polygons. For example, the minimum
weight triangulation can be computed for polygons in cubic time using Klincseks algorithm, but is NP-hard for point sets
[19]. Only a few methods exist for optimal triangulations of point sets. The edge insertion paradigm [2] can be used to optimize several measures in O (n2 log n) or O (n3 ) time (depending on the measure). A triangulation of a point set minimizing
the maximum edge length can be computed in O (n2 ) time [8]. The greedy triangulation, which lexicographically minimizes
the sorted vector of length edges, can be constructed in O (n log n) time [17].
Our problem is more involved, since we aim at optimizing a measure over higher order Delaunay triangulations, therefore
enforcing well-shaped triangles at the same time as optimizing some other measure. Fig. 2 shows an example. There are not
many results on optimal higher order Delaunay triangulations. For the case k = 1, the triangulations have a special structure
that allows a number of measures (for example max or min area triangle, total edge length, number of local minima in a
terrain) to be optimized in O (n log n) time [11]. A few other measures, like minimizing the maximum area ratio of edgeadjacent or vertex-adjacent triangles, can also be optimized eciently [23]. Other measures like maximizing the number of
convex edges or minimizing the maximum vertex degree have been shown to be NP-hard [23]. For k > 1, fewer results are
known. Minimizing local minima in a terrain is NP-hard for orders at least n , where is any positive constant [7].
In this paper we present an extension of the algorithm by Klincsek [16] that allows to optimize a decomposable measure
for a simple polygon over order-k Delaunay triangulations. A straightforward extension of Klincseks algorithm leads to

R.I. Silveira, M. van Kreveld / Computational Geometry 42 (2009) 803813

805

O (kn3 + n3 log n) running time. Our main contribution is improving this to O (k2 n log k + kn log n), by exploiting properties
of this special class of triangulations. This represents a substantial improvement, given that small values of k are most
important [7]. Moreover, we also explain how some other measures that are not decomposable can be optimized following
the same approach, using a more complicated dynamic programming algorithm.
We also explain how to extend our algorithm to triangulate polygons with points inside, and present experimental results
on the structure of order-k Delaunay triangulations that show that in practice, the same approach can be used to triangulate
point sets for small values of k optimally. This constitutes the rst practical result on optimization of higher order Delaunay
triangulations for k > 1.
Note that in this paper we use the standard denition of order of a triangle, as in [11,20], which does not take the
boundary edges of the polygon into account. This implies that for a given polygon and value k, our algorithm may nd
that no order-k triangulation of that polygon exists. In such a case, it is always possible to increase k until a triangulation
is found. The authors recently studied possible denitions of the order of a triangle that take a set of constraining edges
into account [22], which guarantee that a polygon can always be triangulated for any k. However, although seven different
denitions for this notion of constrained order were proposed in [22], no single denition can be regarded as the natural or
right one.
The remainder of this paper is organized as follows. In the next section we introduce higher order Delaunay triangulations more formally, together with some results that will be needed later. Section 3 presents an algorithm to triangulate
polygons optimally, for decomposable measures, using order-k Delaunay triangulations. Some non-decomposable measures
that can also be optimized with a similar method are discussed in Section 4. Section 5 extends the algorithm to be able
to triangulate polygons with points inside. In Section 6 we present experimental results suggesting that the algorithm of
Section 5 can be used to triangulate point sets eciently, due to the special structure of order-k Delaunay triangulations,
for low values of k. Finally, in Section 7 we conclude and present some possible directions for further research.
2. Higher order Delaunay triangulations
In this section we present some basic concepts on higher order Delaunay triangulations, together with some results that
will be needed later. From now on we assume non-degeneracy of the input set P : no four points are cocircular.
Denition 1. (See [11].) A triangle uv w in a point set P is order-k Delaunay if its circumcircle C (u , v , w ) contains at most
k points of P . A triangulation of a set P of points is an order-k Delaunay triangulation if every triangle of the triangulation is
order-k.
Denition 2. (See [11].) For a set P of points, the order of an edge pq between two points p , q P is the minimum number
of points inside any circle that passes through p and q. The useful order of an edge is the lowest order of a triangulation
that includes that edge.
For brevity, we will sometimes write order-k instead of order-k Delaunay, and k-OD edge instead of order-k Delaunay edge.
We also assume that k  1 is a given integer, and write useful edge instead of useful k-OD edge.

Lemma 3. (See [11].) Let uv be a kOD edge, let s1 be the point to the left of vu, such that the circle C (u , s1 , v ) contains no points to

the left of vu. Let s2 be dened similarly but to the right of vu. Edge uv is a useful k-OD edge if and only if uvs1 and uvs2 are k-OD
triangles.
We extend the basic denitions and lemma above with one more lemma.
Lemma 4. Let uv be a useful k-OD edge. There are O (k) order-k triangles that have uv as one of their edges.

Proof. We will show that uv can be part of at most k + 1 triangles on each side. Imagine we slide a circle in contact with u
and v until it touches a rst point r1 to the right of the edge (see Fig. 3). This could potentially be a third point of a triangle

that includes uv because it is possible for the circumcircle of triangle ur1 v to contain less than k + 1 points. If we slide
the circle again until it touches a second point r2 , we now know that the circle contains at least one point (r1 ). Continuing
in this way it can be seen that the circle dened by u, v and the (k + 1)-th touching point, rk+1 , contains at least k points,
hence no further point can be a third point because then the circle would contain k + 1 points. Since an identical argument

can be applied to the left side, we conclude that at most O (k) triangles can have uv as one of its edges. 2
Next we show that all the order-k triangles formed by useful k-OD edges can be computed eciently.
Lemma 5. Let P be a set of n points in the plane. In O (k2 n log k + kn log n) expected time one can compute all order-k triangles of P
that are incident to three useful k-OD edges.

806

R.I. Silveira, M. van Kreveld / Computational Geometry 42 (2009) 803813

Fig. 3. At most k + 1 points can be a third point.

Fig. 4. Finding k-OD triangles incident to uv.

We provide a sketch of the algorithm. An order-k triangle must meet two conditions: its circumcircle must contain at
most k points and it must be empty. For a set of n points there are O (kn) useful k-OD edges [11], and by Lemma 4, a given
useful edge can be part of O (k) order-k triangles. This makes the total number of order-k triangles O (k2 n).
All the useful edges can be computed in O (k2 n + kn log n) expected time [11]. Moreover, without increasing the previous

asymptotic running time, we can store for every useful edge uv, the two sets of points that are contained in the two circles
that determine its usefulness (see Lemma 3). There are at most k of these points on each side. For each side, we will sort
the points according to the order in which they are touched when sliding a circle in contact with u and v (in the same way
as in the proof of Lemma 4). This can be done in O (k log k) time by sorting the centers of the circles.
To nd out which third points can make an order-k triangle, we need to count the number of points inside each circumcircle. This can be done using the two sorted lists of points as follows. We explain how to do it for the right side, the left

side is identical. Let L = {l , . . . , l1 } be the sorted points to the left of uv, and R = {r1 , . . . , r } the ones to its right. See

Fig. 4. The circumcircle of uvr1 , denoted C (u , v , r1 ), contains by denition no points to the right of uv and points to

its left. For the second point, r2 , we know that C (u , v , r2 ) contains exactly one point to the right of uv (namely, r1 ). To nd

out how many points it contains to the left of uv, we check whether l is inside C (u , v , r2 ) or not. If it is, then C (u , v , r2 )

contains exactly points to the left of uv. If it is outside, we go through L until we nd the rst l j that is inside C (u , v , r2 ).

That implies C (u , v , r2 ) contains exactly j points to the left of uv. This is then repeated for r3 , r4 , . . . , r . The running time

is linear in k because both lists are scanned only once, from left to right. Hence, for each useful edge uv we can nd part

of the triangles incident to uv whose circumcircles contain at most k points in O (k log k) time.

Notice that this algorithm, when applied to one edge uv, does not necessarily nd all the order-k triangles adjacent

to uv. It may happen that some of the order-k triangles adjacent to it have a third point that is not among the points

included in the two circles dening the usefulness of uv, because these circles contain at most k points, but may contain
less. However, as the following lemma shows, we are not missing any important triangle.
Lemma 6. Let t = uv w be a non-Delaunay, order-k triangle, such that its three edges are useful order-k. Then there is at least one
edge of t such that one of the circles dening the useful order of that edge contains the third point of t.
Proof. Since t is not a Delaunay triangle, C (u , v , w ) contains at least some point p, which can be in one of three regions
bounded by each of the edges of t and C (u , v , w ). Assume w.l.o.g. the region is the one bounded by uv, and denote it R.
Let q be the point inside R such that C (q, u , v ) contains no other point from R (possibly, q = p). Such a point must be
inside C (u , v , w ), otherwise C (q, u , v ) would contain p. C (q, u , v ) is one of the circles dening the useful order of uv (see
Lemma 3). Since q is inside C (u , v , w ), C (q, u , v ) must contain w, and the result follows. 2
This observation guarantees that, at the end, all the order-k triangles composed of three useful edges, which are the only
ones that can be part of an order-k triangulation, will be taken into account.
Still, some of these triangles may contain points inside, so we need to discard the ones that are not empty. Let uvx
and uv y be two triangles, and let u ( v ) denote the angle of uvx at u (at v), and u ( v ) the same for uv y. It is
easy to see that uvx contains point y if and only if u < u and v < v . Each triangle can be represented by a point in

R.I. Silveira, M. van Kreveld / Computational Geometry 42 (2009) 803813

807

Fig. 5. Dynamic programming method. The value of an optimal triangulation of P 1,6 is stored in L [1, 6].

the plane using its angles at u and at v. The empty triangles are the ones lying on the lower-left staircase of the point set,
and can be found in O (k log k) time by a sweep line algorithm.
The total time needed to nd the triangles for one useful edge is O (k log k). Since there are O (kn) useful edges, all the
useful edges and order-k triangles can be computed in O (k2 n log k + kn log n) expected time, proving Lemma 5.
3. Triangulating polygons
As mentioned in the introduction, Klincseks algorithm allows to triangulate polygons optimally for a large number of
measures using dynamic programming. In this paper we have the additional requirement that the resulting triangulation
must be order-k, therefore the classical algorithm must be adapted to include only order-k triangles.
The input of the algorithm is a polygon P , dened by its vertices in clockwise order: p 0 , p 1 , . . . , pn1 . The output is a
k-OD triangulation of optimum cost.
The dynamic programming algorithm nds an optimal solution by combining solutions of smaller problems in a systematic way. The typical algorithms have O (n3 ) running time and use an n n matrix L, which in our problem has the following
meaning: L [i , j ] is the cost of the optimal k-OD triangulation of the polygon P i , j , dened by the edges of P between p i and
p j , plus edge p j p i . See Fig. 5 for an illustration.
The matrix can be lled in a recursive way. The simplest entries are the ones of the form L [i , i + 1], which have cost 0.
The recursive formula for L [i , j ] is:
L [i , j ] = min

i <q< j

( p i , pq , p j ) L [i , q] L [q, j ]

(1)

The expression ( p i , p q , p j ) denotes the cost of triangle  p i p q p j , and the operator represents a way to combine the
values of the subproblems. Their precise meaning depends on the measure being optimized. Edges p i p q and p q p j must be
diagonals. Triangles that are not contained entirely inside P or are not k-OD have cost +. Checking the latter (verifying
that there are no more than k points inside the circumcircle of the triangle) would take O (log n + k) time [11], but if we
precompute all the order-k triangles for each useful edge (see Section 2) and store them in a perfect hashing table [9], we
can nd out in O (1) time if the triangle is among the order-k triangles.
We can take advantage of the properties of higher order Delaunay triangulation to reduce the running time signicantly.
The main steps of the algorithm are the following (details are given below). In the preprocessing phase we compute all the
useful edges and lter out the ones that are not fully contained inside the polygon. The order-k triangles adjacent to each
useful edge are precomputed. The triangulation algorithm proceeds by applying Eq. (1), using a perfect hashing table to
store the solutions to the subproblems already computed.
For a given edge p i p j , the number of possible third points to form a triangle is not O (n), as in the normal triangulation
problem, but O (k) (see Lemma 4). If for every edge we precompute these O (k) points, we can improve the O (n3 ) dynamic
programming running time to O (kn2 ), after spending O (k2 n log k + kn log n) time in the precomputation of the useful edges
and the order-k triangles (see Lemma 5).
Every time an edge is considered as a candidate to be in an optimal triangulation, we must also check that it does not
intersect the polygon boundary and that it does not lie outside the polygon. This check can be done in O (log n) time per
edge using an algorithm for ray shooting in polygons [13]. This adds an O (kn log n) term to the preprocessing time, which
does not increase the previous asymptotic running time.
Finally, the matrix L has O (n2 ) cells, each corresponding to one potential edge. However, we know that only O (kn)
edges will be useful, so it is not necessary to keep a data structure of quadratic size. In order to avoid wasting time and
space on edges that are not useful, we will not use the standard matrix-based dynamic programming algorithm, but we
will use a memoized version instead. The idea is to use Eq. (1) to compute L [0, n 1], and maintain a perfect hashing
table where we will store all the subproblems already solved. Notice that each subproblem L [i , j ] is associated with the
insertion of an edge p i p j , which must be useful k-OD. Hence, only O (kn) subproblems will be computed and stored. The
same table used to store the order-k triangles incident to an edge can be extended to also store the value of the subproblem
associated with that edge. To solve one particular problem O (k) time is needed, yielding a total running time of O (k2 n),
plus O (k2 n log k + kn log n) preprocessing time.

808

R.I. Silveira, M. van Kreveld / Computational Geometry 42 (2009) 803813

Theorem 7. In O (k2 n log k + kn log n) expected time one can compute an optimal order-k Delaunay triangulation of a simple polygon
with n vertices that optimizes a decomposable measure.
4. Non-decomposable measures
The approach described above can also be used to optimize some other, non-decomposable, measures. The challenging
part is adapting the recursive formula to make the subproblems independent. We show below how this can be done for
minimizing the maximum vertex degree, minimizing the number of local minima (if the points have an elevation), and
optimizing functions of quadrilaterals, such as minimizing the maximum area ratio of triangles sharing an edge.
Interestingly, a similar approach also allows to nd the lowest order completion of a polygon and a set of diagonals,
that is, nding a higher order Delaunay triangulation, with the lowest possible order, which contains the given diagonals.
The more general problem of nding a lowest order completion of a point set with respect to a set of edges is still open;
polynomial time algorithms are known only for k  3 [12].
4.1. Minimizing the maximum vertex degree
When the goal is to minimize the maximum vertex degree, a formula like Eq. (1) must be adapted to take into account
that a vertex can be shared by two different subproblems. A more complicated version of the formula that achieves this is
presented in [14] in the context of triangulating a face of a graph. It can also be used for a polygon, leading to an algorithm
that runs in time O (n4 log n). When the triangulation must be order-k, the adaptations presented in the previous sections
result in an expected running time of O (k2 n2 log n).
Theorem 8. Given a polygon with n vertices, an order-k Delaunay triangulation that minimizes the maximum vertex degree can be
computed in O (k2 n2 log n) expected time.
4.2. Minimizing the number of local minima
A common way to model a terrain is by a TIN (triangulated irregular network), that is, a triangulated point set where
each point has an elevation, and linear interpolation over the triangles is assumed. In many applications of GIS, like terrain analysis for hydrology, local minima, which can be dened as vertices whose neighbors have all higher elevation, are
considered artifacts that should be avoided [18]. An interesting problem is triangulating a terrain while minimizing the
total number of local minima, over order-k Delaunay triangulations. Only heuristics are known for the case k > 1 [7]. Here
we show how to triangulate a polygon while minimizing the number of local minima, assuming that each vertex has an
elevation. This is clearly a less natural problem than the one for point sets, but we believe it is still of interest.
Proceeding as in Section 3, we can insert an edge p i p j and then try all the possible third points p q to create a triangle
 p i pq p j . However, simply solving the two resulting subpolygons recursively can lead to counting the same minimum more
than once, or counting a vertex as a minimum when it cannot be, because it is also connected to some lower vertex that is
not part of the current subpolygon.
To account for this we will extend the recursive formula with two extra parameters as follows: T (i , j , b i , b j ) will denote
the number of local minima of an optimal triangulation of P i j . The two new arguments b i and b j are binary values that determine whether p i and p j can be counted as local minima in the triangulation of the subpolygon. This allows to propagate
this information through the steps of the recursion.
The recursive formula can be adapted by considering for every third point p q three cases, depending upon which point
of the three is the lowest one. Notice that only the lowest vertex of a triangle can be a local minimum. Let l iqj be the index
of the lowest point of { p i , p q , p j }.

liqj = i
T [i , q, bi , 0] + T [q, j , 0, 0]
liqj = j
T [i , q, 0, 0] + T [q, j , 0, b j ]
T [i , j , b i , b j ] = min
i <q< j
min{ T [i , q, 0, 1] + T [q, j , 0, 0], T [i , q, 0, 0] + T [q, j , 1, 0]} liqj = q

The base case is T [i , i + 1, b i , b i +1 ], and has value 1 in two cases: when h( p i ) < h( p i +1 ) and b i = 1, and in the symmetric
case h( p i ) > h( p i +1 ) and b i +1 = 1. In the other cases it has value 0. Notice that, as usual, it must be checked that all the
edges inserted are valid.
The extensions described add only a constant number of computations per edge, hence the asymptotic running time of
the algorithm for general triangulations remains O (n3 ). Adapting the algorithm to work for order-k Delaunay triangulations
can be done following the same approach used in Section 3. Therefore the running time is also the same.
Theorem 9. Given a polygon with n vertices, where each vertex has a height, a triangulation of the polygon that minimizes the number
of local minima can be computed in O (n3 ) time. An order-k Delaunay triangulation that minimizes the number of local minima can be
computed in O (k2 n log k + kn log n) expected time.

R.I. Silveira, M. van Kreveld / Computational Geometry 42 (2009) 803813

809

Fig. 6. Optimizing a function of a quadrilateral. The problem on the left is given by polygon P i , j and a triangle  p i p j pt . To compute S [i , j , t ], each possible
vertex pq is tried and for each choice, two new subproblems arise: S [i , q, j ] and S [q, j , i ] (center and right). Each subproblem is comprised of a subpolygon
(shaded) and one extra triangle (dotted lines).

4.3. Optimizing a function of a quadrilateral


A function that depends on the two triangles incident to one edge (or, equivalently, on a quadrilateral), can also be
optimized for a polygon using dynamic programming. The optimization of this type of measure over higher order Delaunay
triangulations has been studied for point sets for the special case k = 1 [23]. Some of them can be optimized in O (n log n)
time whereas others were shown to be NP-hard. Here we present a recursive formula for any k that allows to triangulate
a polygon while optimizing any min max measure of this type (for example, minimizing the maximum area ratio between
adjacent triangles). The same method works also for optimizing the number of occurrences of some property of a quadrilateral (for example, maximizing the number of convex edges). We explain the idea for a general measure function . As
before, the techniques from Section 3 can be used to constrain the triangulation to be order-k.
The value of an optimal triangulation is dened as:
OPT = min S [i , i 2, i 1]
i

for every i such that ( p i 2 p i 1 p i ) is contained in P (indices wrap around), and where S is recursively dened as:

S [i , j , t ] = min S [i , q, j ] S [q, j , i ] (i , q, j , t )
i <q< j

S [i , j , t ] represents the optimal value of the subpolygon P i , j augmented with an extra xed triangle adjacent to p i p j ,
namely  p i p j pt . Fig. 6 illustrates the idea. The base cases are of the form S [i , i + 2, j ] = (i , i + 1, i + 2, j ), and correspond
to the value of a triangle on the boundary of the polygon plus one extra triangle adjacent to it. The main difference
with Eq. (1) is that now subproblems are not just subpolygons, but subpolygons with an attached triangle (external to the
subpolygon).
Since the table needed for this algorithm has cubic size, the running time for general triangulations is O (n4 ).
For order-k Delaunay triangulation, the approach used in the previous sections reduces the number of subproblems to
be stored to O (k2 n), and solving one problem, if the required subproblems are already computed, takes O (k) time, yielding
a triangulation time of O (k3 n). The preprocessing time is the same as in Section 3, therefore the overall expected running
time is O (k3 n + kn log n).
Theorem 10. Given a polygon with n vertices, a triangulation of the polygon that optimizes a min max function of edge-adjacent
triangles, or one that optimizes the number of occurrences of a property of edge-adjacent triangles, can be computed in O (n4 ) time. An
order-k Delaunay triangulation that optimizes such a function can be computed in O (k3 n + kn log n) expected time.
5. Triangulating polygons with points inside
In this section we consider the more general problem of nding an optimal triangulation of a simple polygon that
contains h components in its interior. A component can be either a point or a connected component made of several points
connected by edges. We will denote the polygon with the components by P . We can reuse the algorithm from the previous
section if we connect one vertex of each component to some other vertex in order to remove all the loose parts. To nd
the optimal triangulation we must try, in principle, all the possible ways to make these connections. The number of them
depends on h and on the order k. In principle, there are O (n) possible k-O D edges that connect a vertex of a component
to the outer boundary, even for constant k. However, we can reduce the number of edges that we need to consider to O (k).
Lemma 11. Let P be a polygon with h components inside. There is a collection of O (k)h sets, of h edges each, such that: (i) for every set
in the collection, the edges in the set connect all the components in P to the outer boundary; (ii) any order-k Delaunay triangulation
includes the edges of some set in the collection.
Proof. First we show that for a given component in P , any order-k triangulation T of P must connect the component to the
rest of the polygon by one of O (k) edges. Let u be the topmost point among the boundary points of the components inside
the polygon. Every vertex above u is part of the outer polygon boundary. Let uv be an edge of the Delaunay triangulation

810

R.I. Silveira, M. van Kreveld / Computational Geometry 42 (2009) 803813

of the point set induced by P and its components (ignoring the edges), with v higher than u. Since uv is a Delaunay edge,
we know from [11] that the useful k-OD edges that cross it have O (k) endpoints on each side of uv. If uv is not part of
T , at least one of the useful edges that cross it must be. Let xy be the rst of these edges (the rst one encountered when
going from u to v along uv) in T , then triangle uxy must be part of T . This implies that edges ux and u y are part of it
as well. Hence, either uv or one of the O (k) possible edges of type uz (for z = x or z = y) must be in T , and connects v to
a higher point of the outer boundary of P .
Following the same idea, for each of the h components we can nd a set of O (k) useful k-OD edges such that any
triangulation T connects each component using one of these O (k) edges. The result follows. 2
Using the previous result, our algorithm will try the O (k)h different ways to connect the loose components in P . Let
P 1 , . . . , P be the O (k)h different polygons that are tried, and let H i be the set of new boundary edges of P i . For each P i ,
besides computing the boundary, we must compute the intersections between the O (kn) useful edges and the new h edges
in H i , which were added to connect the loose components. This is because during the triangulation we need to consider
only edges that make the polygon simply-connected.
The computation of these intersections can be done once and maintained between successive polygons without increasing the asymptotic running time. During the preprocessing phase, we will compute all the intersections between useful k-OD
edges. A useful k-OD edge can intersect O (k2 ) other useful k-OD edges [11]. Therefore the total number of intersections is
O (k3 n), and they can be computed in time O (kn log kn + k3 n)=O (kn log n + k3 n) [4]. We store for each useful edge all the
other useful edges that it intersects and in addition we keep a counter. The counter will be used to keep track of how many
edges in H i each useful edge intersects.
The algorithm will iterate through the polygons in such a way that two consecutive polygons P i and P i +1 differ only in
the edge chosen for one of the components. Then during the (i + 1)-th step the counters for H i are already computed, and
one can compute the counters for H i +1 very easily, as follows: let e out be the edge that is removed and e in the new edge
(that is, H i +1 = H i \{e out } {e in }). Firstly, all the O (k2 ) useful edges that intersect e out must have their counters decreased by
one. Secondly, all counters of the O (k2 ) edges that intersect e in are incremented by one. Both sets of edges were previously
computed during the preprocessing phase and can be accessed in constant time. Hence the time needed to update the
intersection information from one polygon to the next one is O (k2 ).
We conclude that the total time required to compute all the new intersections is O (kn log n + k3 n) for the preprocessing
and O (k2 ) per polygon. Note that the useful edges and order-k triangles do not need to be recomputed, since they only
depend on the point set, which has not changed.
Triangulating each generated polygon using the algorithm from the previous section takes O (k2 n) time, yielding a total
time of O (k2 n log k + nk log n + kn log n + k3 n) + O (k)h (k2 + k2 n)=O (kn log n) + O (k)h+2 n (because h  1).
Theorem 12. An optimal order-k Delaunay triangulation of a simple polygon with n boundary vertices and h  1 components inside
that optimizes a decomposable measure can be computed in O (kn log n) + O (k)h+2 n expected time.
6. Application to point sets
Any point set can be optimally triangulated using the results of Section 5 if it is seen as a polygon made of its convex
hull with points inside. In general, this will lead to a running time exponential in n, so this is of no practical use. For low
order Delaunay triangulations, the situation is better. Given a point set and an order k, there are xed edges that are present
in any order-k triangulation, and partition the convex hull of the point set into a number of polygons with components
inside (Fig. 7). For k = 1, it is known that these polygons are always empty triangles or quadrilaterals [11], which simplies
the optimization of several measures. As k increases, the number of xed edges decreases until it is reduced to little more
than the convex hull. Moreover, for k > 1 the polygons may be larger and may have many components inside. However,
our experiments on the structure of higher order Delaunay triangulations suggest that in practice, for small values of k, the
appearance of such polygons is rather unlikely.
We summarize part of the results of experiments carried out on randomly generated point sets, which show that for
small values of k, the polygons created contain only a few components. The experiments consisted in generating random
point sets of between 1000 and 5000 points, and for different values of k, computing the partition into polygons with
components inside given by the xed edges. The size of the polygons and the number of components was registered. These
are the two factors, besides k, involved in the running time of the algorithm of Section 5. Tables 1 and 2 show the results

Fig. 7. Fixed edges: k = 1 (left), k = 2 (center) and k = 4 (right).

R.I. Silveira, M. van Kreveld / Computational Geometry 42 (2009) 803813

811

Table 1
Structure of order-k triangulations: average/maximum number of components per polygon for random point sets of n points, averaged over 200 runs.
k

n = 1000

n = 2000

n = 3000

n = 4000

n = 5000

1
2
3
4
5
6
7
8
9
10

0.00/0.00
0.00/0.00
0.00/0.03
0.00/0.44
0.01/1.15
0.05/2.22
0.18/5.52
0.55/18.95
1.45/61.05
3.08/115.55

0.00/0.00
0.00/0.00
0.00/0.03
0.00/0.68
0.01/1.35
0.05/2.63
0.17/6.77
0.53/24.19
1.47/108.70
3.33/225.36

0.00/0.00
0.00/0.00
0.00/0.04
0.00/0.80
0.01/1.50
0.04/2.77
0.16/6.82
0.52/29.85
1.47/150.83
3.44/332.13

0.00/0.00
0.00/0.00
0.00/0.06
0.00/0.94
0.01/1.53
0.04/2.91
0.16/7.38
0.52/31.28
1.52/205.30
3.62/443.35

0.00/0.00
0.00/0.00
0.00/0.07
0.00/0.90
0.01/1.57
0.04/3.13
0.16/7.87
0.52/35.85
1.53/242.75
3.72/552.41

Table 2
Structure of order-k triangulations: average/maximum size of polygons for random point sets of n points, averaged over 200 runs. Since the dynamic
programming algorithm works by considering triangles, the size is measured as the number of Delaunay triangles that the polygon contains.
k

n = 1000

n = 2000

n = 3000

n = 4000

n = 5000

1
2
3
4
5
6
7
8
9
10

1.35/2.00
1.99/6.82
2.84/12.77
4.02/23.54
5.80/43.97
8.63/91.66
13.48/222.25
21.71/569.40
34.99/1294.76
49.91/1755.54

1.35/2.00
1.99/7.41
2.85/13.84
4.04/25.95
5.82/52.90
8.68/108.55
13.56/263.83
22.40/749.15
37.26/2425.46
56.27/3566.14

1.35/2.00
1.99/7.54
2.84/14.60
4.04/28.02
5.83/54.55
8.74/115.63
13.72/292.36
22.76/980.10
38.32/3434.07
59.04/5341.94

1.36/2.0
2.00/7.63
2.85/15.14
4.04/28.37
5.84/56.25
8.74/123.63
13.83/310.12
23.21/1038.78
39.89/4716.74
62.56/7195.49

1.35/2.00
2.00/7.77
2.85/15.59
4.04/29.95
5.85/59.72
8.78/129.58
13.86/337.01
23.41/1223.43
40.58/5661.38
64.77/9034.72

for point sets between 1000 and 5000 points, and k = 1, . . . , 10. It is worth mentioning that since the convex hull of the
point set limits the growth of the polygons, the results may be inuenced slightly by boundary effects.
Some examples of the polygons found are shown in Fig. 8. The images show only the xed edges, for each value of k.
Each image was chosen after 100 runs as the one containing a polygon with the largest number of components inside. That
polygon is highlighted in red. It can be seen that for k = 5, 6 the largest polygon seems local, whereas for k = 7, 8 it extends
over a large part of the point set.
A few observations are in order. The experiments conrm that for k  3, it is very unlikely to nd polygons with components inside. Even though for k  2 one can build examples where that is the case, they hardly arise in random point sets.
Even for orders up to 5 or 6, the size of the polygons and number of components are small enough to be useful for practical
purposes. As a result, nding optimal triangulations that in general are NP-hard, like the minimum weight triangulation,
can be done in practice if the Delaunay order is low enough, as long as the measure is decomposable. The small values of k
are the most useful in practice, for several reasons. On the one hand, as k increases, the shape of the triangles deteriorates.
On the other hand, previous experimental results [7], related to realistic terrain modeling, have shown that low values of k
are enough to obtain important improvements on several terrain measures (like the number of local minima), making small
values of k particularly interesting for these applications.
7. Conclusions and future work
We studied algorithms to nd higher order Delaunay triangulations of polygons that optimize a decomposable measure.
Based on an existing technique for polygon triangulation, we proposed an algorithm to compute an optimal triangulation
of a polygon restricted to order-k triangulations. Their specic properties allowed us to reduce an O (n2 ) factor to O (k2 ), a
substantial improvement since k will be, in general, much smaller than n [7]. We also explained how this method can be
extended to optimize some non-decomposable measures like number of local minima, or functions of quadrilaterals like the
area ratio of incident triangles.
For the more general problem of triangulating a polygon with components inside optimally, we presented an algorithm
that is xed-parameter tractable for k = O (1).
We also gave experimental evidence suggesting that the specic structure of order-k Delaunay triangulations, for small
values of k, makes the algorithm presented here applicable to point sets. This constitutes the rst practical result on optimal
higher order Delaunay triangulations for k > 1, allowing to optimize any decomposable function over a class of well-shaped
triangulations.
There are several topics for further research. A rst one is studying what other measures, relevant to specic applications, can be optimized with a similar approach. A second interesting topic is combining the algorithms presented here
with some denition of higher order constrained Delaunay triangulations [22]. We would like to have a denition of the
order of a triangle that is inuenced by the polygon edges, allowing us to always obtain some triangulation that is close

812

R.I. Silveira, M. van Kreveld / Computational Geometry 42 (2009) 803813

Fig. 8. Examples of the subdivision induced by xed edges for random point sets of 1000 points, and different values of k. For each example the polygon
with the largest number of components is highlighted in red. (For interpretation of the references to color in this gure legend, the reader is referred to
the web version of this article.)

to the constrained Delaunay triangulation. As mentioned in the introduction, there seems to be no single denition of the
constrained order of a triangle that generalizes the concept of constrained Delaunay triangulation to higher orders in a completely natural way. The authors concluded in [22] that the choice of such a denition will depend on the application in
mind. The algorithms presented in this paper are general, in the sense that they were designed to be able to work for a
large number of measures. Even though any of the denitions proposed in [22] can easily be plugged in into the algorithms,
many of the improvements on the running time achieved here would be lost. It may be interesting to try to adapt the
polygon triangulation algorithms to some of the denitions proposed in [22].
Acknowledgements
We thank Remco Burema for carrying out the experimental work in this paper.
References
[1] M. Benkert, J. Gudmundsson, H.J. Haverkort, A. Wolff, Constructing interference-minimal networks, in: Proc. 32nd Int. Conf. on Current Trends in Theory
and Practice of Computer Science (SOFSEM06), 2006, pp. 166176.
[2] M. Bern, H. Edelsbrunner, D. Eppstein, S. Mitchell, T.S. Tan, Edge insertion for optimal triangulations, Discrete Comput. Geom. 10 (1) (1993) 4765.
[3] M. Bern, D. Eppstein, Mesh generation and optimal triangulation, in: D.-Z. Du, F.K. Hwang (Eds.), Computing in Euclidean Geometry, second ed., in:
Lecture Notes Series on Computing, vol. 4, World Scientic, Singapore, 1995, pp. 47123.
[4] B. Chazelle, H. Edelsbrunner, An optimal algorithm for intersecting line segments in the plane, J. ACM 39 (1) (1992) 154.
[5] L.P. Chew, Constrained Delaunay triangulations, Algorithmica 4 (1989) 97108.
[6] L. De Floriani, B. Falcidieno, C. Pienovi, Delaunay-based representation of surfaces dened over arbitrarily shaped domains, Comput. Vision Graph.
Image Process. 32 (1985) 127140.
[7] T. de Kok, M. van Kreveld, M. Ler, Generating realistic terrains with higher-order Delaunay triangulations, Comput. Geom. Theory Appl. 36 (2007)
5265.
[8] H. Edelsbrunner, T.S. Tan, A quadratic time algorithm for the minmax length triangulation, SIAM J. Comput. 22 (1993) 527551.
[9] M.L. Fredman, J. Komlos, E. Szemeredi, Storing a sparse table with O (1) worst case access time, J. ACM 31 (3) (July 1984) 538544.
[10] P.D. Gilbert, New results in planar triangulations, Report R-850, Coordinated Sci. Lab., Univ. Illinois, Urbana, IL, 1979.
[11] J. Gudmundsson, M. Hammar, M. van Kreveld, Higher order Delaunay triangulations, Comput. Geom. Theory Appl. 23 (2002) 8598.
[12] J. Gudmundsson, H. Haverkort, M. van Kreveld, Constrained higher order Delaunay triangulations, Comput. Geom. Theory Appl. 30 (2005) 271277.

R.I. Silveira, M. van Kreveld / Computational Geometry 42 (2009) 803813

[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]

813

J. Hershberger, S. Suri, A pedestrian approach to ray shooting: Shoot a ray, take a walk, J. Algorithms 18 (1995) 403431.
G. Kant, Augmenting outerplanar graphs, J. Algorithms 21 (1996) 125.
J.M. Keil, T.S. Vassilev, Algorithms for optimal area triangulations of a convex polygon, Comput. Geom. Theory Appl. 35 (3) (2006) 173187.
G.T. Klincsek, Minimal triangulations of polygonal domains, Discrete Math. 9 (1980) 121123.
C. Levcopoulos, D. Krznaric, The greedy triangulation can be computed from the Delaunay triangulation in linear time, Comput. Geom. Theory Appl. 14
(1999) 197220.
D. Mark, Network models in geomorphology, in: M.G. Anderson (Ed.), Modelling Geomorphological Systems, John Wiley & Sons, 1988, pp. 7397,
Chapter 4.
W. Mulzer, G. Rote, Minimum weight triangulation is NP-hard, in: Proc. 22nd Annu. ACM Sympos. Comput. Geom., 2006, pp. 110.
M. Neamtu, Delaunay congurations and multivariate splines: a generalization of a result of B.N. Delaunay, Trans. Amer. Math. Soc. 359 (7) (2007)
29933004.
P.P. Pebay, T.J. Baker, A comparison of triangle quality measures, in: Proceedings of the 10th International Meshing Roundtable, 2001, pp. 327340.
R.I. Silveira, M. van Kreveld, Towards a denition of higher order constrained Delaunay triangulations, Comput. Geom. Theory Appl. 42 (2009) 322337.
M. van Kreveld, M. Ler, R.I. Silveira, Optimization for rst order Delaunay triangulations, in: F. Dehne, J.-R. Sack, N. Zeh (Eds.), WADS 2007, in: LNCS,
vol. 4619, Springer, Heidelberg, 2007, pp. 175187.

You might also like