Parameterized Complexity of Geometric Problems
Transcript of Parameterized Complexity of Geometric Problems
Parameterized Complexity of
Geometric Problems
PANOS GIANNOPOULOS1,*, CHRISTIAN KNAUER2 AND SUE WHITESIDES3
1Institut fur Informatik, Humboldt-Universitat zu Berlin, Unter den Linden 6, 10099 Berlin, Germany2Institut fur Informatik, Freie Universitat Berlin, Takustraße 9, D–14195 Berlin, Germany
3School of Computer Science, McGill University, 3480 University Street, Montreal, Canada H3A 2A7
*Corresponding author: [email protected]
This paper surveys parameterized complexity results for hard geometric algorithmic problems.
It includes fixed-parameter tractable problems in graph drawing, geometric graphs, geometric
covering and several other areas, together with an overview of the algorithmic techniques used.
Fixed-parameter intractability results are surveyed as well. Finally, we give some directions for
future research.
Keywords: Computational geometry, graph drawing, Parameterized algorithms
Received 5 February 2007; revised 31 March 2007
1. INTRODUCTION
This paper surveys parameterized complexity results for
NP-hard geometric problems. Geometric problems arise fre-
quently in application domains as diverse as computer
graphics [1], computer vision [2–4], VLSI design [5],
geographic information systems [6, 7], graph drawing [8]
and robotics [9, 10], and typically involve (sets of) geometric
objects, such as, points, line segments, balls or polytopes,
which usually lie in some metric space. Designing efficient
algorithms and data structures for such problems is the main
focus in the field of computational geometry [11–13].
Many geometric problems can be formulated as combina-
torial optimization problems where the objective is to maxi-
mize or minimize a function subject to constraints induced
by a given collection of geometric objects; by exploiting the
geometric nature of such problems, one can obtain faster
and simpler algorithms. Geometric optimization has been a
topic of extensive research, and many general techniques
have been developed that yield efficient polynomial-time
algorithms for a wide range of problems (see the excellent
survey by Agarwal and Sharir [14]).
1.1. Polynomial-time approximation schemes
In order to cope with NP-hard geometric optimization
problems, researchers have, so far, mainly resorted to the tra-
ditional toolkit of approximation algorithms. A polynomial-
time approximation scheme (PTAS) for a minimization
problem is an algorithm, or rather a sequence of algorithms,
which, given any e . 0 and any problem instance of size n,
produces a solution for the problem with cost at most (1 þ
e) times the cost of an optimum solution in O(nf(1/e)) time,
where f is an arbitrary computable function (for maximization
problems, the cost of the solution produced is at least (1 2 e)
times that of an optimum solution). A PTAS is called effici-
ent (EPTAS), if it is uniformly polynomial, i.e., it runs in
O(f(1/e) . nc) time for every e, where c is a constant indepen-
dent of e. The ultimate approximation scheme is an EPTAS
where f is polynomial; such a scheme is called fully PTAS
(FPTAS).
PTASs have been found for many important geometric pro-
blems (see the surveys by Bern and Eppstein [15] and Arora
[16]). While some of these schemes are efficient [Euclidean
traveling salesman problem (TSP) problem [17], minimum
k-median problem [18]], others exhibit prohibitively high-
running times even for moderate error values [19]. Moreover,
there are also quite a few problems for which no PTAS is
known, e.g. minimum weight triangulation (MWT) [20], or
likely to exist (under standard complexity theoretic assump-
tions), e.g. minimum k-center problem (when k is part of the
input) [21, 22].
1.2. Parameterized complexity
Parameterized complexity theory offers an alternative frame-
work for the study of hard algorithmic problems by measuring
their complexity in terms of one or more parameters, explicitly
or implicitly given by their underlying structure, in addition to
THE COMPUTER JOURNAL, 2007
# The Author 2007. Published by Oxford University Press on behalf of The British Computer Society. All rights reserved.For Permissions, please email: [email protected]
doi:10.1093/comjnl/bxm053
The Computer Journal Advance Access published July 27, 2007
the problem input size. The main idea is to devise exponential-
time exact—as opposed to approximation algorithms for
NP-hard problems, confining exponentiality to the parameters.
In concrete applications, such parameters are hoped to take
relatively small values, thus, resulting in affordable algori-
thms. For an introduction to the field of parameterized com-
plexity theory, we refer the reader to the recent textbooks by
Flum and Grohe [23] and Niedermeier [24], in addition to
the first, classic work of Downey and Fellows [25]. We
review some basic definitions here.
A problem with input instance of size n and with a non-
negative integer parameter k is fixed-parameter tractable
(FPT) if it can be solved by an algorithm that runs in
O(f(k) . nc) time, where f is a computable function depending
only on k and c is a constant independent of k. The class of
all FPT problems is the parameterized complexity analog
of P. Parameterized complexity theory provides a set of
general algorithmic techniques for proving fixed-parameter
tractability, ranging from bounded search trees and kerneliza-
tion methods to color coding and dynamic programming on
tree decompositions. Such techniques have successfully been
used for a variety of parameterized algorithmic problems in
graph theory, logic and computational biology.
Parameterized complexity theory also provides a frame-
work for establishing fixed-parameter intractability. For this
purpose, a whole new (infinite) hierarchy of complexity
classes has been introduced, the ‘W-hierarchy’, with FPT
being its lowest class, and hardness is sought via suitable
reductions, which preserve fixed-parameter tractability
between parameterized problems. Hardness for some level
of the hierarchy can be thought as the parameterized complex-
ity analog of NP-hardness, and, as in classical complexity
theory, intractability results are conditional and, thus, serve
as relative lower bounds. The working assumption for para-
meterized complexity is that all levels of the W-hierarchy
are pairwise distinct. For example, a problem that is
W[1]-hard for some parameterization is not FPT for this para-
meterization unless FPT ¼ W[1].
Every optimization problem has a standard parameteriza-
tion, where the parameter is just an upper or lower bound on
the value of the objective function, that is, it measures the
size or the quality of the solution given by the objective func-
tion. If an optimization problem has an EPTAS, then its stan-
dard parameterization is in FPT [26–28]. Hence, by showing
that the standard parameterization of an optimization problem
is hard for some level of the W-hierarchy, we provide strong
evidence that this optimization problem does not admit an
EPTAS. The survey by Marx [29] discusses many more
aspects of the interplay between the theories of parameterized
complexity and approximation algorithms. While most
tractability and intractability results involve standard parame-
terizations of optimization problems, one strength of parame-
terized complexity theory lies in its ability to accommodate a
variety of other types of problems as well as different
parameterizations. For example, finding an optimal vertex
cover in a graph is in FPT when parameterized by the size
of the cover (standard parameterization), and also when para-
meterized by the tree width of the graph.
Geometric problems often involve a variety of natural para-
meters that measure certain properties of the input objects.
However, only a small number of NP-hard geometric pro-
blems have been studied within the framework of parameter-
ized complexity theory so far. The field is still in its infancy
and the results are rather scattered. We believe that this is
mainly due to different current research foci of the commu-
nities involved, rather than the limited applicability of para-
meterized complexity to geometric problems. This is well
illustrated with the example of the MWT problem (see also
Section 2.5.1): given a set of n points in the plane, one asks
for a triangulation with the minimum total length of edges.
This is a well-known problem in computational geometry
whose complexity status was unknown for more than three
decades, until it was recently proved to be NP-hard and not
to admit an FPTAS [30]. As already mentioned earlier, no
PTAS is known for this problem. However, the problem is
FPT when parameterized with the number k of points in the
interior of the convex hull of the input points. The first
known such algorithm runs in O(6kn5 log n) time [31]; succes-
sive improvements brought the running time down to O(4kkn4)
[32], O(2kkn3þ n3) [33] and O(2c
ffiffiffikp
log kk2n3) [34].
This paper is organized as follows. Section 2 presents FPT
geometric problems, while Section 3 presents intractability
results. We conclude with a few general research directions
in Section 4.
2. FPT GEOMETRIC PROBLEMS
2.1. A retrospective
Using parameters in order to achieve faster algorithms for geo-
metric problems that are nevertheless in P has always been a
common practice. Typical parameters for this purpose
include the number of (non)convex corners, the diameter of
a point set or the minimum inter-point distance in a point
set, and measures of how ‘slim’ or ‘fat’ the objects are. In a
few cases, special attention has been given to the actual depen-
dence on parameters that can take small values in practice.
One such famous example is Megiddo’s algorithm [35] for
linear programming in Rd with n constraints, where d, i.e.
the number of variables, is the parameter, and which runs in
O(22d
n) time1; the running time was later substantially
improved to dO(d)n by Dyer and Frieze [36], Chazelle and
Matousek [37], and to O(d2n þ eO(ffiffiffidp
In dÞ) (randomized) by
1When d is bounded, linear programming can be solved in strongly poly-
nomial time, i.e. time that depends only on n and not on the binary represen-
tation of the input; when d is unbounded, only weakly polynomial time
algorithms are known.
Page 2 of 13 P. GIANNOPOULOS
THE COMPUTER JOURNAL, 2007
Matousek et al. [38] using some quite advanced techniques
from computational geometry.
What about NP-hard geometric problems? After identifying
a problem as such, much effort is often put into getting fast
approximations, while many times little or nothing is said
about solving the problem exactly even by the trivial brute
force exponential algorithm. In the course of writing this
survey, we were able to spot a couple of early algorithms
which can now be seen as being in FPT.
One such interesting example is the following. Deineko and
Woeginger [39] gave an O(kln2)-time dynamic programming-
based algorithm for the convex-hull-and-l-line Euclidean TSP,
where k points lie on l almost parallel line segments in the
interior of the convex hull of a total of n points, such that
the supporting lines of all these segments intersect the hull
into two common edges. Thus, this problem is in FPT when
parameterized with respect to the number of points k and the
number of lines l. This result has been recalled much later
by Deineko et al. [40] who again used dynamic programming
to give an O(2kk2n)-time algorithm for a more general TSP
problem where k is the number of points in the interior of
the convex hull, but with no restriction on their positions
(see Section 2.5.3).
There are several NP-hard problems where important
classes of input instances are directly characterized by a par-
ticular parameter. For some of these problems, researchers
have been able to give algorithms whose running time is
some polynomial the degree of which is independent of the
parameter, which is considered a constant. However, some-
times such constants are not mentioned explicitly, and some-
times they can be of rather ‘astronomical’ value. This
situation is well depicted in the following two problems
drawn from computational topology; such problems arise fre-
quently in applications where one needs to examine topo-
logical as well as geometric properties of the input objects.
Mohar [41] presented a linear time algorithm for the follow-
ing problem: given an arbitrary fixed surface S, i.e. of fixed
genus, and a graph G, can G be embedded in S? The algori-
thm assumes the standard RAM with uniform cost measure
(for a discussion on models of computation see the next para-
graph), and thus, the problem is FPT with respect to the
genus of the surface S. Unfortunately, the actual dependence
on the genus is nowhere to be seen. On the other hand,
Chambers et al. [42] recently showed that the problem of
finding the shortest splitting cycle on a combinatorial
surface is not only NP-hard, but also FPT with respect to
the genus of the surface. The dependence on the genus is
given and the algorithm runs in gO(g)n2 log n time (see
Section 2.6). This is probably the first time that ‘fixed-
parameter tractability’ appeared in an article in a major
computational geometry conference.
It is reasonable to believe that many more FPT problems
could be found scattered in the literature. Even if the algori-
thms seem to be trivial, one can gain valuable insight by
asking the following questions: Which properties made the
problem tractable? Which parameters were used? Were the
results obtained using techniques that are now considered
standard in parameterized complexity, or by some different
approach? In the latter case, can the results be improved by
using the standard techniques? At this point the reader
should be cautious: many results in computational geometry
assume real RAM, a much stronger model of computation
than the standard RAM adopted in parameterized complexity
theory. Extending parameterized complexity theory to accom-
modate other models of computation is a current research issue
and one should be careful before claiming (in)tractability of
geometric problems.
2.2. Independent set on geometric intersection graphs
We start with a few definitions that will be needed in other
sections as well. A topological graph G(V, E) is a graph
embedded in the plane such that any two edges intersect in
at most a finite number of points. A geometric graph is a
topological graph where the edges are straight-line segments.
The intersection graph of a set of geometric objects S is a
graph G(V, E), with a one-to-one mapping from S to V, and
with two vertices being connected if and only if the corres-
ponding two objects intersect. When embedded in the
plane, such a graph need not necessarily be a geometric
graph.
2.2.1. Independent set on disk graphs
Consider the following problem: given a collection of n disks
in the plane with bounded radius ratio and l-precision, i.e. all
disks have radius in the range [1, s] and all centers are pair-
wise at least l . 0 apart, is there a set of k non-intersecting
disks? Basically, the bounded radius property says that no
disk has arbitrarily small radius, while the l-precision prop-
erty asserts that no two centers are arbitrarily close. For
fixed s and l, Alber and Fiala [43] gave an 2O(ffiffiffikp
)þ
nO(1)-time algorithm for this problem. Their algorithm is
based on kernelization.
2.2.1.1. Kernelization. The main idea of kernelization is
to pre-process the input instance in order to reduce its size
by removing its ‘easy’ parts before actually trying to solve
the problem. This can lead to considerable speedups. Of
course this idea is nothing new, but kernelization systematizes
pre-processing (see the survey by Niedermeier [44]). For-
mally, let L be a parameterized problem, i.e. L consists of
pairs (I, k), where I is the problem instance and k is the para-
meter. A polynomial time computable function K: (I, k)!
(I0, k0) is a problem kernel reduction if there is a computable
function g: N! N such that (i) (I, k) [ L if and only if (I0,
k0) [ L, and (ii) jI0j� g(k0) and k0 � k. The instance (I0, k0) is
called the problem kernel with g(k0) being its size. It has
PARAMETERIZED COMPLEXITY OF GEOMETRIC PROBLEMS Page 3 of 13
THE COMPUTER JOURNAL, 2007
been shown that a parameterized problem is in FPT if and only
if it admits a reduction to a problem kernel.
THEOREM 2.1. Independent set on disk graphs with
bounded radius ratio s and l-precision admits a reduction
to a linear problem kernel.
Proof. For a collection of disks D ¼ fD1, . . . , Dng in the
plane, let GD ¼ (VD, ED) be the intersection graph of D, i.e.
VD ¼ fv1, . . . ,vng and ED ¼ fvivjjDi > Dj = ;g. A disk
graph GD is given together with its representation, i.e. the col-
lection of disksD, and a natural embedding in the plane with vi
lying at the center of Di (see Fig. 1). Also, let m(GD) ¼
m(D) ¼ m(< D) be the space covered by the union of disks
in D, where m( . ) is the standard Lebesgue measure.
First, the problem is reduced to a ‘geometric’ problem
kernel, where the size of the kernel is bounded by O(k),
when measured by m( . ). This is done by a simple volume
argument, which uses the property of bounded radius ratio.
First, if m(D) . 9ps2k, then a greedy algorithm can easily
select k pairwise disjoint disks, since the disks representing
the neighbors of any vertex v [ VD can occupy an area of at
most 9ps2. Second, when m(D) , pk, there is no set of k dis-
joint disks since any such set needs space of at least pk. Other-
wise, pk � m(GD) � 9p s2k and the input graph GD is
basically the kernel. It can be computed in nO(1) time, which
is needed to compute m(GD).
Next, using the l-precision property, the geometric problem
kernel is converted to a proper problem kernel by bounding the
size of its vertex set jVDj by m(GD). This is done in the follow-
ing way: when l � 2 then we have that m(GD) � (p/4)l2jVDj
since the disks with radius l/2 around all centers of disks in Dare pairwise disjoint and contained in the area covered by the
original disks; when l . 2 then of course m(GD) � pjVDj. In
both cases, this leads to a problem kernel of size ck, where c is
a constant depending on s and/or l. A
Since the problem admits a problem kernel, it is in FPT:
trivially checking all possible k subsets of the kernel for inde-
pendence leads to an O(kO(k)þ nO(1))-time algorithm. A faster
algorithm can be obtained using graph separators.
2.2.1.2. Graph separators. A vertex separator VS # V of
an undirected graph G(V, E) separates V into two parts VA,
VB for which (i) VA < VS < VB ¼ V and (ii) no edge in E con-
nects a vertex of VA to a vertex of VB. The triple (VA, VS, VB) is
called a separation of G. For a function f : N! Rþ and a
graph class G, an f( . )-vertex separator theorem on G for
the measure j . j and some constants a , 1, b . 0 is a
theorem asserting that for any G [ G there is a separation
(VA, VS, VB) of G such that (i) jG[VS]j � b . f(jGj) and (ii)
jG[VA]j, jG[VB]j � a . jGj, where G[X] denotes the subgraph
of G induced by the set X # V.
Alber and Fiala [43] gave affiffi�p
-separator theorem for the
counting measure (with constants a ¼ 3/4 and b ¼ O(s/l))
on the class of disk graphs with bounded radius ratio and
l-precision. Moreover, Alber et al. [45] proved that the para-
meterized independent set problem (as well as other ‘glue-
able’ graph problems) on a graph class G that admits affiffi�p
-separator theorem for the counting measure can be
solved in O(2ffiffiffikp
þ TK (n, k)) if it has a linear size problem
kernel that can be constructed in TK (n, k) time. Combining
these two ideas with the problem kernel described earlier
leads to an O(2ffiffiffikp
þ nO(1))-time algorithm.
2.2.2. Independent set on segment graphs
Marx [46] showed that the independent set problem on
segment graphs is in FPT when parameterized with the
number of the different directions of the segments: given a
set of n segments with at most d different directions, one
can decide whether there are k pairwise disjoint segments in
2O(k2d2 log d) n log n time. For the same problem, Kara and
Kratochvıl [47] gave a faster algorithm, which runs in
O(d2kk9dkþ2þ n2
þ k3dmn) time, where m is the number of
the edges of the segment graph, and which works also
in the case where only the intersection graph is given (not
the segments themselves). Both algorithms are based on
kernelization.
2.3. Graph drawing problems
We begin with an illustrative example showing the application
of kernelization and bounded search tree techniques to
a problem in layered graph drawing, namely, 2-layer
planarization.
A 2-layer drawing of a bipartite graph is one that positions
the vertices on two parallel lines and draws edges as straight
line segments between endpoints. The endpoints of an edge
must lie on different layers. A graph is said to be biplanar if
it has a 2-layer drawing without edge crossings, and a subset
E0(G) of the edges of a graph G is said to be a biplanarizing
set if its removal from G results in a biplanar graph.
The 2-layer planarization problem is as follows: given a
graph G (not necessarily bipartite) and a positive integer k,
does G have a biplanarizing set of size at most k?FIGURE 1. A disk graph with a straight-line embedding in the
plane.
Page 4 of 13 P. GIANNOPOULOS
THE COMPUTER JOURNAL, 2007
For the related 1-layer planarization problem, an ordered
assignment of vertices to the top layer is given, and one asks
if it is possible to remove at most k edges so that the vertices
on the second layer can be ordered so that no edges cross.
Dujmovic et al. [48] showed that both problems are in FPT
and gave an O(3kjGj)-time algorithm for the 1-layer planariza-
tion problem and an O(k6kþ jGj)-time one for the 2-layer pla-
narization problem. We describe the main ideas of their
2-layer planarization algorithm in the following four sections.
2.3.1. Forbidden subgraphs
Clearly, cycles cannot be drawn on two layers without edge
crossings. Another graph with this property is the 2-claw.
This consists of a vertex v of degree 3 (the center of the
2-claw), to which are attached three neighbors of degree 2;
each of these in turn is joined to a leaf node. Thus, there are
six nodes in addition to the center, v. It is easy to see that if
a graph contains a 2-claw as a subgraph (not necessarily
induced), then it cannot be drawn on two layers without
edge crossings (see Fig. 2).
In fact, it can be shown that the absence of cycles and
2-claws characterizes biplanar graphs: G is biplanar if and
only if it contains no cycles and no 2-claws.
A wreath of G is a connected component consisting of a
cycle with some leaves attached. A pendant wreath of G is
similar to a wreath, except that it contains a unique ‘attach-
ment’ vertex w that is connected to one or more non-leaf ver-
tices of G that do not lie on the cycle of the pendant wreath.
Thus, w is a cut vertex for G.
First, as a prelude to kernelization, wreaths and pendant
wreaths will be destroyed by choosing certain edges from
them to go into a biplanarizing set. This will reduce the
initial budget, k. Then, the remaining graph will be shrunk
to a kernel, without affecting the presence of 2-claws and
cycles.
2.3.2. Destroying wreaths and pendant wreaths
To destroy wreaths and pendant wreaths, one can choose a set
of edges, one from each wreath and pendant wreath, for
removal from G. It can be shown that it is possible to do
this so that only one edge is removed from each wreath or
pendant wreath, and so that the set of edges removed is con-
tained in at least one biplanarizing set for G of minimum
size. (A pendant wreath whose cycle has three nodes will con-
tribute the cycle edge not incident to its attachment vertex; a
pendant wreath whose cycle has length .3 will contribute a
cycle edge adjacent to its attachment vertex.) With these
edges removed from G, and the budget k reduced accordingly
to k0, one proceeds to the kernelization step, knowing that the
graph no longer contains wreaths or pendant wreaths.
2.3.3. Shrinking the graph to a kernel
A convenient fact for generating a kernel for this problem is
the following: there is always a biplanarization set of
minimum size that contains no leaf edges (we omit the
proof). This means that we can first simplify the graph by
replacing all the leaf edges at a vertex v of degree .1 by a
single leaf edge. This single edge may belong to a 2-claw. If
so, the original graph may have had many 2-claws, containing
various of the leaf edges attached to v. However, according to
the fact above, there is a biplanarization set of minimum size
that destroys all such 2-claws by removal of some non-leaf
edge. Thus, retaining a single leaf edge at v represents the
possibility that v may belong to a 2-claw in which v is adjacent
to the center.
Once the leaf edges are amalgamated in this way, one pro-
ceeds to shrinking long caterpillar paths. A caterpillar path is
an induced subgraph consisting of a path with leaf edges pos-
sibly attached to some interior nodes of the path. One or both
endpoints of the path, but no intermediate path nodes, may
be joined to other nodes of the graph. Note that an internal
node on a caterpillar path cannot serve as the center of a
2-claw, so 2-claws can only touch nodes within distance 2
from the end of a caterpillar path. For this reason, leaves
attached to interior points at distance 2 or more from the
path endpoints can be removed without affecting the presence
of 2-claws. Finally, the path, stripped of these leaves, can be
shortened without affecting the presence of 2-claws and
cycles (see Fig. 3).
Finally, it can be shown that either this reduced graph does
not have a biplanarizing set of size within the remaining
budget, in which case an algorithm looking for a biplanarizing
set can terminate, or it has at most 20k edges, in which case it
may be worthwhile to continue the search. Any small biplanari-
zing set for the reduced graph (together with the edges chosen
to destroy wreath components and pendant wreaths) points
out a corresponding biplanarizing set for the original graph.
At this point, one could simply test each of the 20kk0
� �sets of
k0 edges to see if one of them is a biplanarizing set for GFIGURE 2. Two forbidden subgraphs of biplanar graphs: a 2-claw
and a cycle.
PARAMETERIZED COMPLEXITY OF GEOMETRIC PROBLEMS Page 5 of 13
THE COMPUTER JOURNAL, 2007
(here, k0 is k reduced by the number of edges chosen to destroy
wreaths and pendant wreaths). With more care, one can obtain
an algorithm based on these kernelization ideas having
running time in O(ffiffiffikp
. 17kþjGj).
2.3.4. Bounded search tree approach
Further improvement is possible by applying a bounded search
tree approach to the shrunken, kernel graph. Rather than
testing the sets of k0 edges in a brute force manner, one
creates a search tree whose nodes correspond to subgraphs
of the kernel graph, together with associated budgets
reduced by the node’s depth in the search tree. At each internal
tree node, one looks for a vertex v of the subgraph of the kernel
graph associated with this tree node such that v has at least
three non-leaf neighbors. Then exploring the current subgraph
of the kernel graph from this vertex, one easily finds either a
2-claw with center v or a cycle of length 3 or 4 containing v
(see Fig. 4).
Let C denote the edges of the 2-claw or cycle found. Now,
one creates jCj children for the search tree node, one for each
of the edges in C, since this cycle or 2-claw must be destroyed
in order to biplanarize the graph. The graph at a child node is
the graph at the parent, with an edge of C removed, and the
budget at the child node is one less than that for the parent.
This search tree can have depth no greater than k0 � k,
where k0 is the budget for the kernel graph. Since each internal
node has at most six children, this gives a search tree of size at
most 6k and a running time in O(k . 6kþ jGj).
THEOREM 2.2. Given a graph G and an integer k . 0,
the 2-layer planarization problem on G can be solved in
O(k . 6kþ jGj) time.
2.3.5. Experimental results
Experiments of Suderman and Whitesides [49] with an
implementation based on the approach given above, after
further refinements and additions, gave promising results.
The implementation could find optimum solutions provided
the biplanarizing set had size at most about k ¼ 18, compar-
able to results based on integer linear programming appro-
aches. The running time was reduced to O(6kþ jGj) by
detecting a forbidden subgraph (if one exists) at each node
in the search tree in constant time.
2.3.6. 1- and 2-layer minimization
The 2-layer minimization problem is defined as follows: given
a bipartite graph G, and an integer k, does G admit a 2-layer
drawing with at most k edge crossings? Again, if a permutation
of the vertices in one layer is fixed, then the problem becomes
that of 1-layer minimization. The latter can be solved in
O(1.4656kþ kn2) time [50], an improvement over the earlier
O(fkn2)-time algorithm by Dujmovic and Whitesides [51],
where f is the golden ratio. For the 2-layer minimization
problem, the fastest algorithm runs in O(232(2þ2k)3
n) time [52].
2.3.7. h-layer problems
The 1- and 2- layer problems can be viewed as special cases of
h-layer problems. A well-known approach for drawing certain
types of graphs, particularly directed acyclic graphs, is to
assign each node to one of h layers, and then to seek permu-
tations of the nodes within each layer that reduce the
number of crossing points (giving a layered crossing minimiza-
tion problem), or that attribute the crossings to a small number
of edges whose removal leaves the drawing crossing free
(giving a layered planarization problem). In particular, the
‘Sugiyama method’, due to Sugiyama et al. [53], fixes a per-
mutation of the nodes assigned to the top layer, then seeks a
good permutation for layer 2 with respect to layer 1, etc. It
sweeps up and down the h layers, adjusting the permutation
of the current layer at each step. While node assignment to
FIGURE 4. A vertex that has three non-leaf neighbors is either the
center of a 2-claw or lies on a cycle of length 3 or 4.
FIGURE 3. Shrinking a long path while preserving 2-claws and
cycles.
Page 6 of 13 P. GIANNOPOULOS
THE COMPUTER JOURNAL, 2007
layers is typically done in a separate, preceding step, it is of
course possible to ask for the layer assignments and the per-
mutations within each layer to be specified all at once.
Dujmovic et al. [52] showed that various h-layer problems
are in FPT, where h is an additional parameter. The running
times are in O(232(hþ2k)3
n) and are rather impractical.
2.3.8. General crossing minimization
The crossing number of a graph is the minimum possible
number of edge crossings when the graph is drawn nicely in
the plane. Here the edges need not be drawn straight, but the
drawing must satisfy the usual, expected properties. For
example, for each pair of distinct edges of the graph, the corres-
ponding curves must intersect in at most one interior point.
In answer to a question from the book by Downey and
Fellows, Grohe [54] found an algorithm with running time
in O(f(k) . n2) to test whether the input graph has crossing
number at most k.
2.4. Configurations with few crossings in geometricgraphs
Knauer et al. [55] studied the following problem: given a
graph G ¼ (V, E) together with an embedding E in the plane,
compute a certain subgraph of G with the minimum number
of crossings in the embedding E. A crossing is defined as a
pair of edges that intersect at a common point. They con-
sidered subgraphs such as spanning trees, fixed source and
target paths (s-t paths), cycles, matchings (with given cardi-
nality), and k-factors, i.e. k-regular spanning subgraphs for
k [ f1, 2g. They showed that all these problems are in FPT
when parameterized with the number of crossings in the
embedding E of the input graph G.
To illustrate the approach, we show that the problem of
deciding if G has a spanning tree without crossings (in the
embedding E) is in FPT when parameterized with the
number of crossings k in the embedding E: since G has a
crossing-free spanning tree if and only if it has a crossing-free
planar spanning graph it is sufficient to examine all maximal
crossing-free subgraphs of G and test if one of them is
connected. This can be done as follows.
(i) Remove all edges X that participate in a crossing from
G. The resulting graph G0 is crossing-free.
(ii) For all maximal crossing-free subsets, H of edges from
X check if the graph G0 < H is connected.
It can easily be shown that the number of maximal crossing-
free subsets of edges from X is 2k, so the algorithm runs in
O((m þ n) þ 2kn) time.
Note that this problem admits a kernel of size O(k): if G0 has
more than k þ 1 connected components, we can immediately
conclude that we cannot add k edges to make it connected,
so in the second step we can actually shrink the connected
components of G0 to a single vertex and therefore only have
to consider a graph G00 with O(k) edges and vertices in the
connectivity test.Again G00 < X has at most 2k crossing-free
subgraphs and so the algorithm runs in O((m þ n) þ k2k) time.
2.5. Triangulations, convex partitions, TSP, and
‘distance from triviality’
This section deals with optimization problems on planar point
sets, parameterized with the the number of inner points, i.e.
points in the interior of the convex hull. This is a quite
novel parameterization, which follows the general idea of
‘distance from triviality’ [56]: some problems are solvable
in polynomial time when the points are in convex position,
and hence, the number of inner points as a parameter inter-
polates from triviality when there is no inner point to intract-
ability in the general case.
2.5.1. Minimum weight triangulation
An MWT of a point set is a triangulation with the minimum
total length of edges. The problem was only recently shown
to be NP-hard by Mulzer and Rote [30].
For a set P of n points in the plane, with k inner points,
Hoffmann and Okamoto [31] gave an O(6kn5 log n)-time
algorithm to compute an MWT of P. This approach was
later improved by Spillner [33] and Borgelt et al. [32] to run
in time O(2kkn3þ n3) and O(4kkn4), respectively. A further
improvement was recently obtained by Knauer and Spillner
[34], whose algorithm runs in O(2cffiffiffikp
log kk2n3) time and
uses techniques based on small graph separators.
The underlying approach of all these algorithms is simple:
they perform an exhaustive search among all triangulations
of P and keep track of the one with minimum total length
encountered so far. We will now illustrate the basic idea
common to all these algorithms in some more detail.
A partition of P is a set of non-crossing line segments with
endpoints in P (these segments are called edges); all partitions
contain the edges from the boundary of the convex hull C(P)
of P. Such a set of edges E induces a partition R(E) of C(P)\E
into finitely many maximal connected regions. A triangulation
of P is a partition where all the regions are empty triangles (see
Fig. 5 for an example of a partition of a set of points into three
regions).
The search starts with an arbitrary edge e of the convex hull
of P. For every triangulation of P, there is a unique triangle t
that has e in the boundary (see Fig. 6a for an example). The
algorithm now branches on the set of empty triangles with ver-
tices in P and e in the boundary. The edges of C(P) together
with the edges of t form a partition E of P.
In general, we have a partition E of P and for one of the
regions in E a particular triangulation is fixed. This region is
called fixed and all other regions open. The goal is to
compute MWTs for all open regions in E.
The algorithm maintains the following invariant: every
open region R is a topological disk with possibly some
PARAMETERIZED COMPLEXITY OF GEOMETRIC PROBLEMS Page 7 of 13
THE COMPUTER JOURNAL, 2007
points of P in its interior, and the part of the boundary of R that
is contained in the interior of the convex hull C(P) of P is con-
nected. Figure 6b shows an example after several branching
steps.
In order to find the MWT of an open region R (while main-
taining the invariant), we select an edge e that is part of the
boundary of R but not an edge of C(P). Then we branch on
the set of empty triangles with vertices in P that are contained
in R and that have edge e in the boundary.
So, the open regions that we encounter during our search for
an MWT can be viewed as the subproblems we have to deal
with. Note that an open region is defined by the part of its
boundary that is contained in the interior of C(P). This part
is connected and therefore forms a simple (possibly closed)
polygonal path; all vertices of this polygonal path are inner
points, except at most two that are outer points.
Since the exhaustive search repeatedly runs into the same sub-
problems, one can employ dynamic programming, i.e. store
MWTs for subproblems, once computed, in a table where they
can be looked up (in constant time). The total number of sub-
problems encountered by the algorithm is bounded by the
number of polygonal paths that define the open regions encoun-
tered as subproblems. By a result of Sharir and Welzl [57], the
number of crossing-free polygonal paths on l � k vertices is
O(dl) for some constant d , 87. Thus, the number of subpro-
blems can be bounded by O(88k n2). A subproblem can easily
be processed as outlined above: after selecting edge e we have
to consider at most n triangles. For each of these triangles, we
must check whether the triangle is empty. This can easily be
done in O(k) time for each triangle. Hence, we can process a sub-
problem in O(kn) time and thus obtain a running time of
O(88k kn3).
Hoffmann and Okamoto [31] obtained a much faster algori-
thm result based on the observation that it is sufficient to con-
sider subproblems that are defined by polygonal paths that are
monotone with respect to the x-axis: the number of these
x-monotone paths is in O(2k n2). Employing some additional
arguments of a geometric flavor Spillner [33] and Borgelt
et al. [32] improved the running time to O(2kkn3þ n3) and
O(4kkn4), respectively.
In order to give a better upper bound on the number of ver-
tices of the polygonal paths that define subproblems, Knauer
and Spillner [34] used techniques based on small graph separa-
tors. Since for MWT no reduction to a problem kernel is
known, the general approach by Alber and Fiala [43] that
reduces the input to a linear size problem kernel and then
applies the separator result to this problem kernel, (cf.,
Section 2.2) does not work here. The main result of Knauer
and Spillner is that in the dynamic programming algorithm
for MWT outlined above, it is sufficient to consider only
subproblems defined by crossing-free polygonal paths with
O(ffiffiffikp
) vertices.
THEOREM 2.3. Given a set P of points in the plane, k
of which in the interior of the convex hull of P, an MWT
of P can be found in O(2cffiffiffikp
log kk2n3) time.
2.5.2. Minimum (weight) convex partition
A convex partition of a planar point set P is a set of non-
crossing straight line segments (edges) with endpoints in P
that partition its convex hull into a set of empty convex
regions. A region is empty if it contains no points of P in its
interior. In the minimum convex partition problem (MCP),
the task is to compute a convex partition having a minimum
number of regions, whereas in the minimum weight convex
partition problem (MWCP), the task is to compute a convex
partition of minimum total edge length.
For the MCP problem, Hoffmann and Okamoto [31] and
Spillner [58] gave algorithms that run in O(k6k25 216kn) and
O(2kk3n3þ n log n) time, respectively; k is again the
number of points in the interior of the convex hull of P.
For the MWCP problem, Spillner also gave an algorithm
with the same time bound as above, which is actually better
than the more recent one of O(213k k4k28n3) by Borgelt
et al. [32] for the same problem.
2.5.3. Euclidean TSP
For the planar Euclidean TSP on n points, with the number of
inner points k as a parameter, Deineko et al. [40] gave an
algorithm that runs in O(k!kn) time and O(k) space, and one
that runs in O(2kk2n) time and O(2kkn) space.
FIGURE 5. Partition of a point set into the set R(E)¼ fR1, R2, R3g of
three regions.
FIGURE 6. Exhaustively searching for an MWT.
Page 8 of 13 P. GIANNOPOULOS
THE COMPUTER JOURNAL, 2007
2.6. Graphs on surfaces
2.6.1. Shortest splitting cycle on a surface
Let M be a compact, connected and orientable surface without
boundary. A combinatorial surface is an abstract surface M
together with a weighted undirected graph G(M), embedded
on M so that it cuts M into topological disks. A path on M is
only allowed to be a walk in G, and its length is the sum of
the weights of the edges traversed by the path, counted with
multiplicity. A cycle on M is called a splitting cycle if it is
simple, non-contractible and separates M. The complexity of
a combinatorial surface G(M) is the total number of vertices,
edges, and faces of G. For an introduction to topology of
surfaces, see, for example, Kinsey [59]. For more details on
combinatorial topology and topological graph theory, see
Gross and Tucker [60], Mohar and Thomassen [61], and
Hatcher [62].
Chambers et al. [42] proved that finding a shortest splitting
cycle on a combinatorial surface is NP-hard and gave an algori-
thm that computes a shortest splitting cycle in gO(g)n2 log n
time, where g is the genus of the surface and n is the comple-
xity of the surface.
2.6.2. Embedding a graph on a surface
Given an arbitrary fixed surface S, i.e. of fixed genus, and a
graph G, can G be embedded in S? Mohar [41] showed
that this problem is in FPT with respect to the genus of the
surface by giving an algorithm that runs in time linear
in the size of the graph. However, the actual dependence on
the genus is not given. It would be interesting to determine
the exact running time of the algorithm, with the genus con-
sidered as a parameter, and to see whether it can be improved.
2.7. Geometric covering
Langerman and Morin [63] studied an abstract NP-hard cover-
ing problem that models a number of concrete geometric (as
well as purely combinatorial) covering problems. For the
sake of simplicity we will not describe their result in its full
generality. Instead, we illustrate the underlying ideas with
the following simple geometric covering problem: decide if
a set P of n points in the plane can be covered by k lines.
The problem is in FPT when parameterized with k. To see
this, consider the following bounded search tree algorithm:
every node of the tree contains k bins (each bin corresponds
to a line), and each bin has room for two points. A bin is
called open if it contains less than two points, and closed
otherwise. Each closed bin defines a unique line (the one
through the bin’s points) and therefore covers some (at least
two) of the points of P. In the root of the search tree, all
bins are open. The search processes the points of P in some
fixed order. To advance the search down the tree from a
node v one proceeds as follows: the children of v correspond
to all possible ways of assigning the next point of P to the
open bins of v. If we take the next point of P and put it in
an open bin b, we create a new child u of v. If b is now
closed in u, all remaining points from P that are covered by
b are removed. If all bins of a vertex are closed, then the
search stops. In that case it can be checked in O(kn) time if
the bins cover all the points. The depth of this search tree is
2k, and the degree of each vertex is at most k, so the total
size of the tree is O(k2k), and therefore the algorithm runs in
O(k2kþ1n) time.
The problem admits a kernel of quadratic size: to this end,
we consider the set L of O(n2) lines defined by all pairs of
points in P and the set F # L of these lines that contain
more than k þ 1 points. Each line of F has to be part of each
cover of P with at most k lines, so if jFj . k, then the points
in P cannot be covered with k lines. If jFj � k we delete the
points of P covered by the lines of F. We call the resulting
point set P0 and we set k0 ¼ k 2 jFj. Each line in L\F covers
at most k points of P0, so if jP0j .kk0, the set P0 cannot be
covered with k0 lines (and thus P cannot be covered with k
lines). In case jP0j� kk0 � k2, we can check if P0 can be
covered with k0 lines in O(k2kþ2) time using the search tree
method sketched above. The total runtime of the algorithm
is therefore O(n2þ k2kþ2).
The problems considered in [63] are parameterized with the
number of sets allowed in the covering, denoted by k, and the
combinatorial dimension of the problem, denoted by d. For
the geometric problems (as above), the latter parameter is
directly related to the dimension of the underlying Euclidean
space. Among others, the following two geometric covering
problems fit into the general framework.
2.7.1. Covering points with hyperplanes and its dual
problem
Given a set of n points in Rd, is there a set of k hyperplanes
such that each point of S lies on at least one hyperplane?
The dual problem is the following: given a set of n hyper-
planes in Rd, is there a set of k points such that each hyper-
plane in S contains at least one of the points? In both
problems, the combinatorial dimension equals the geometric
dimension d.
2.7.2. Covering points with spheres. Given a set of n
points in Rd, is there a set of k hyperspheres such that each
point in S lies on the surface of at least one of the hyper-
spheres? In this problem, the combinatorial dimension is
equal to d þ 1. Note that this problem becomes W[1]-hard
when the points can lie in the interior of the spheres (see
Section 3.3.2).
Both problems can be solved with a deterministic algorithm
that runs in O(kdkn) time and a randomized Monte–Carlo one
that runs in O(kd(kþ1)þ c2dkd(dþ1)/2eb(dþ1)/2cn log n) time and
succeeds with probability at least 12n2c, where c . 0 is
some constant.
PARAMETERIZED COMPLEXITY OF GEOMETRIC PROBLEMS Page 9 of 13
THE COMPUTER JOURNAL, 2007
3. FIXED-PARAMETER INTRACTABLEGEOMETRIC PROBLEMS
This section reviews hardness results for parameterized geo-
metric problems. Most of these results are based on reductions
from the standard parameterization of the clique problem, or
k-clique, which is W[1]-complete [64]: given a graph G
and a parameter k . 0, is there a clique of size k in G?
Of course, k-independent set on general graphs is also
W[1]-complete [64].
3.1. Visibility independence and domination problems
Consider the following visibility problem: given a set S of line
segments in the plane and a set P of points, decide if there is a
set U , P of k points that mutually cannot see each other
when visibility is blocked by the segments in S. That is, if
Gvis ¼ (P, Evis) denotes the visibility graph of P with respect
to S, i.e. fp, q g [ Evis if and only if pq > S ¼ ;, we ask if
Gvis contains an independent set of size k.
This problem is easily seen to be W[1]-hard since each
graph G ¼ (V, E) can be realized as a visibility graph Gvis.
Assume that V ¼ f1, . . . , ng and let P ¼ fp1, . . . , png be the
vertices of a regular n-gon.
We surround each pi by a small axis-aligned square Bi (con-
sisting of four segments) that initially blocks all visibility to
the remaining points. For each edge fi, j g [ E, we remove
a small segment bij from the boundary of Bi and a small
segment bji from the boundary of Bj such that pipj >(Bi < Bj) ¼ ;, i.e. pi and pj can see each other now. The seg-
ments bij and bji are chosen in such a way that prps > (bij <bji) ¼ ; for all r, s [ f1, . . . , ng\fi, jg, i.e. removing these seg-
ments does not allow any other new pair of points to see each
other. The set S consists of the segments forming the bound-
aries of the Bis after all edges have been considered (see
Fig. 7 for an illustration of the construction).
It can be shown that the reduction can be computed in poly-
nomial time. This implies that the independence problem in
visibility graphs is W[1]-hard.
Using the same approach and reducing from k-dominating
set in general graphs, which is W[2]-complete [65], one can
show that the following problem is W[2]-hard: decide if
there is a set D , P of k points that see all the points in P\D
(again, visibility is blocked by the segments in S), i.e. we
ask if Gvis contains a dominating set of size k.
Note that if the coordinates of P and the endpoints of S are
rational numbers, the problems are clearly in W[1] and W[2],
respectively.
3.2. Geometric intersection graphs
3.2.1. Independent set on intersection graphs
Finding an independent set of size k in the intersection graph
of unit disks and axis-parallel unit squares in the plane is
W[1]-hard [66]. The reduction is from k-clique. Note that the
problem is in W[1], and hence W[1]-complete, since the
same problem is W[1]-complete for general graphs. Marx
[46] has recently generalized his reductions and showed that
the independent set problem on intersection graphs of unit
line segments with an arbitrary number of directions is
W[1]-complete as well; note that, as mentioned in Section
2.2.2, the problem is in FPT when parameterized with the
maximum number of directions.
3.2.2. Dominating set on intersection graphs
Given an intersection graph G, the task is to find a set S of k
vertices such that each vertex of the graph is either in S or is
a neighbor of a member of S. Marx [46] proved the following:
dominating set is W[1]-hard for intersection graphs of axis-
parallel unit squares, axis parallel line segments and unit
disks in the plane. As mentioned above, dominating set is
W[2]-complete for general graphs, but is in W[1], and hence
W[1]-complete, for axis-parallel unit square graphs and axis-
parallel line segment graphs [46]. It is an open problem
whether the problem is in W[1] for unit disk graphs as well.
3.3. Other problems
3.3.1. Robot motion planning
Let a (rational) polyhedron be specified by a finite union of
convex polyhedra in three-dimensional Euclidean space,
each defined by a finite set of linear inequalities with rational
coefficients. The three-dimensional Euclidean generalized
mover’s problem can be defined as follows.
Input: A finite set O of obstacle rational polyhedra, an
object to be moved given as a finite set P of rational polyhedra
which are freely linked at distinguished linkage vertices such
that P has k degrees of freedom of movement, and initial and
final rational positions pI and pF of P in three-dimensional
Euclidean Space.
Question: Is there a legal movement of P from pI to pF, i.e. is
there a continuous sequence of simultaneous translations and
rotations of the polyhedra of P such that at each point in
time each polyhedron of P intersects no obstacle of O and
furthermore it intersects no other polyhedron of P except at
its specified linkage vertices?FIGURE 7. Realization of an abstract graph as a visibility graph.
The lines of sight (dashed) correspond to the edges.
Page 10 of 13 P. GIANNOPOULOS
THE COMPUTER JOURNAL, 2007
Reif [67] proved that the three-dimensional Euclidean gene-
ralized mover’s problem is PSPACE-hard. Cesati and
Wareham [68] adapted Reif’s proof and showed that the
problem is W[SAT]-hard when parameterized with the
number of degrees of freedom k.
3.3.2. Geometric covering
Deciding whether a given set P of points in the plane can be
covered by k unit squares is W[1]-hard [66]. The reduction
is from k-clique; according to D. Marx (Personal Communi-
cation), the reduction carries over to covering with unit
disks in the plane as well.
We observe here that this result holds also for the ‘dual’
problem of k-piercing [69]: for a given set S of squares in
the plane, decide whether there exists a set of k-piercing
points, such that every square of S is intersected by at least
one of the k points. This can be easily seen via the following
standard reduction from the k-covering to the k-piercing
problem: place a unit square around each point of P, such
that the square is centered at the point. This set of squares
can be pierced by k points if and only if P can be covered
by k unit disks. This is actually an FPT-reduction, and hence
k-piercing is W[1]-hard as well.
3.3.3. Shape matching
Given two point sets A, B [ Rd, with jAj ¼ m , jBj ¼ n,
deciding whether A is congruent, i.e. equal up to translations
and rotations, to a subset of B (point subset congruence
problem) is W[1]-hard with respect to the dimension d [70].
The reduction is from k-clique, and the parameter d is linear
in the size of the clique k. The latter implies that an O(mno(d))-
time algorithm for this problem exists only if k-clique can be
solved in O(no(k))-time, which in turn implies that SNP ,DTIME(2o(n)) [71]. This example shows that the actual depen-
dence of the parameters in a parameterized reduction is
important for getting tight lower bounds.
The above hardness result generalizes to any pointset dis-
tance D for which D(A, B) ¼ 0$A , B; this is a desired prop-
erty for any distance that is used to find small patterns in larger
ones, e.g. directed Hausdorff distance, Earth mover’s distance.
Since mini D(i(A), B) ¼ 0 ,i0(A) , B for some isometry i0,
the problem of minimizing D under isometries, when d is
part of the input, is not in FPT, unless FPT ¼ W[1].
4. RESEARCH DIRECTIONS
We have reviewed parameterized complexity results for geo-
metric problems. There are numerous issues open for further
research, several of which are summarized in the following.
(i) Investigate the parameterized complexity of high-
dimensional geometric problems with respect to their
dimension. This approach could prove to be useful
for proving lower bounds even for problems with no
direct formulation as decision problems, e.g. nearest
neighbors search problems [72].
(ii) Investigate the parameterized complexity of compu-
tational topology problems with respect to the genus
of the input surface; see, for example, Chambers
et al. [42] for a few open problems of this type.
(iii) Find alternative parameterizations based on the
‘distance from triviality’ approach, especially for
geometric problems for which no PTAS is known, as
it was done for the MWT problem.
(iv) Are there parameterized geometric problems that have
an implicit graph structure where the tree width can be
used a parameter?
(v) Investigate the efficiency of suitable FPT algorithms
for geometric problems by performing extensive
experiments.
ACKNOWLEDGEMENTS
We would like to thank Mike Fellows for his support, com-
ments and ideas in the course of writing this survey.
REFERENCES
[1] Dobkin, D. and Teller, S. (2004) Computer Graphics. In
Goodman, J. and O’Rourke, J. (eds), Handbook of Discrete
and Computational Geometry, (2nd edn.), pp. 1095–1116.
CRC Press LLC.
[2] Alt, H. and Guibas, L. (1999) Discrete Geometric Shapes:
matching, Interpolation, and Approximation. In Sack, J.
and Urrutia, J. (eds), Handbook of Computational Geometry,
pp. 121–153. Elsevier Science.
[3] Guibas, L. (2004) Modeling Motion. In Goodman, J.
and O’Rourke, J. (eds), Handbook of Discrete and
Computational Geometry, (2nd edn), pp. 1117–1134. CRC
Press LLC.
[4] O’Rourke, J. and Toussaint, G. (2004) Pattern recognition. In
Goodman, J. and O’Rourke, J. (eds), Handbook of Discrete
and Computational Geometry, (2nd edn), pp. 1135–1162.
CRC Press LLC.
[5] Sato, M. and Ohtshuki, T. (1987) Applications of computational
geometry to vlsi layout pattern design. Integr. VLSI J., 5, 303–
317.
[6] van Kreveld, M. (2004) Geographic information systems. In
Goodman, J. and O’Rourke, J. (eds), Handbook of Discrete
and Computational Geometry, (2nd edn), pp. 1293–1314.
CRC Press LLC.
[7] Floriani, L.D., Puppo, E. and Magillo, P. (1999) Applications of
computational geometry to geographical information systems.
In Sack, J. and Urrutia, J. (eds), Handbook of Computational
Geometry, pp. 333–388. Elsevier Science.
[8] Tamassia, R. and Liota, G. (2004) Graph drawing. In
Goodman, J. and O’Rourke, J. (eds), Handbook of Discrete
and Computational Geometry, (2nd edn), pp. 1163–1187.
CRC Press LLC.
PARAMETERIZED COMPLEXITY OF GEOMETRIC PROBLEMS Page 11 of 13
THE COMPUTER JOURNAL, 2007
[9] Sharir, M. (2004) Algorithmic Motion Planning. In Goodman, J.
and O’Rourke, J. (eds), Handbook of Discrete and
Computational Geometry, (2nd edn), pp. 1037–1064. CRC
Press LLC.
[10] Halperin, D., Kavraki, L. and Latombe, J.-C. (2004) Robotics.
In Goodman, J. and O’Rourke, J. (eds), Handbook of Discrete
and Computational Geometry, (2nd edn), pp. 1065–1093.
CRC Press LLC.
[11] Preparata, N. and Shamos, M. (1990) Computational Geometry:
An Introduction. (2nd edn). Springer-Verlag.
[12] Mulmuley, K. (1994) Computational Geometry: An
Introduction Through Randomized Algorithms. Prentice Hall.
[13] de Berg, M., van Kreveld, M., Overmars, M. and
Schwarzkopf, O. Computational Geometry: Algorithms and
Applications (2nd edn). Springer-Verlag.
[14] Agarwal, P. and Sharir, M. (1998) Efficient algorithms for
geometric optimization. ACM Comput. Surv., 30, 412–458.
[15] Bern, M. and Eppstein, D. (1996) Approximation Algorithms
for Geometric Problems. In Hochbaum, D. (ed),
Approximation Algorithms for NP-Hard Problems. PWS
Publishing.
[16] Arora, S. (2002) Approximation schemes for NP-hard
geometric optimization problems: a survey. http://www.cs.
princeton.edu/~arora/pubs/arorageo.ps.
[17] Rao, S. and Smith, W. (1998) Approximating Geometric
Graphs via ‘spanners’ and ‘banyans’. Proc. 30th Annu. ACM
Sympos. Theory of Computing (STOC), Dallas, TX, pp. 540–
550. ACM.
[18] Har-Peled, S. and Mazumdar, S. (2004) On coresets for k-means
and k-median clustering. Proc. 36th Annu. ACM Sympos. Theory
of Computing (STOC), Chicago, IL, pp. 291–300. ACM.
[19] Downey, R. (2003) Parameterized complexity for the skeptic.
Proc. 18th IEEE Ann. Conf. Computational Complexity,
Aarhus, Denmark, pp. 147–169.
[20] Remy, J. and Steger, A. (2006) A Quasi-polynomial time
Approximation Scheme for Minimum Weight Triangulation.
Proc. 38th Annu. ACM Sympos. Theory of Computing
(STOC ), Seattle, WA, pp. 316–325.
[21] Gonzalez, T. (1985) Clustering to minimize the maximum
intercluster distance. Theor. Comput. Sci., 38, 293–306.
[22] Feder, T. and Greene, D. (1988) Optimal Algorithms for
Approximate Clustering. Proc. 20th Annu. ACM Sympos.
Theory of Computing (STOC), Chicago, IL, pp. 434–444. ACM.
[23] Flum, J. and Grohe, M. (2006) Parameterized Complexity
Theory. Texts in Theoretical Computer Science. An EATCS
Series, Vol. XIV. Springer-Verlag.
[24] Niedermeier, R. (2006) Invitation to Fixed-Parameter
Algorithms. Oxford Lecture Series in Mathematics and its
Applications. Oxford University Press.
[25] Downey, R. and Fellows, M. (1999) Parameterized Complexity.
Springer-Verlag.
[26] Bazgan, C. (1995) Scheemas d’approximation et complexitae
parametree. Technical report, Universite Paris Sud.
[27] Cai, L. and Chen, J. (1997) On fixed-parameter tractability and
approximability of NP optimization problems. J. Comput. Syst.
Sc., 54, 465–474.
[28] Cesati, M. and Trevisan, L. (1997) On the efficiency of
polynomial time approximation schemes. Inf. Process. Lett.,
64, 165–171.
[29] Marx, D. Parameterized complexity and approximation,
algorithms. Comput. J. 10.1093/comjnl/bxm048.
[30] Mulzer, W. and Rote, G. (2006) Minimum-Weight Triangulation
is NP-hard. Proc. 22nd ACM Sympp. on Computational
Geometry (SoCG). Sedona, AZ, pp. 1–10. ACM.
[31] Hoffmann, M. and Okamoto, Y. (2004) The Minimum
Weight Triangulation Problem with Few Inner Points. Proc.
1st Int. Workshop Parameterized and Exact Computation
(IWPEC ), Lecture Notes in Computer Science, Vol. 3162,
200–212.
[32] Borgelt, C., Grantson, M. and Levcopoulos, C. (2006) Fixed
parameter algorithms for minimum weight partitions. Proc.
22nd European Workshop Computational Geometry EWCG,
pp. 5–12.
[33] Spillner, A. (2005) A Faster Algorithm for the Minimum
Weight Triangulation Problem with few Inner Points. Proc.
1st ACiD, Durham, UK, pp. 135–146.
[34] Knauer, C. and Spillner, A. (2006) A Fixed-Parameter
Algorithm for Minimum Weight Triangulation based on Small
Graph Separators. Proc. 32nd Int. Workshop on Graph
Theoretic Concepts in Computer Science, Lecture Notes in
Computer Science, vol. 4271, 49–57.
[35] Megiddo, N. (1984) Linear programming in linear time when
the dimension is fixed. J. ACM, 31, 114–127.
[36] Dyer, M. and Frieze, A. (1989) A randomized algorithms for
fixed-dimension linear programming. Math. Program., 44,
203–212.
[37] Chazelle, B. and Matoussek, J. (1996) On linear-time
deterministic algorithms for optimization problems in fixed
dimension. J. Algorithms, 21, 579–597.
[38] Matousek, J., Sharir, M. and Welzl, E. (1996) A subexponential
bound for linear programming. Algorithmica, 16, 498–516.
[39] Deineko, V. and Woeginger, G. (1996) The
convex-hull-and-k-line traveling salesman problem. Inf.
Process. Lett., 59, 295–301.
[40] Deineko, V., Hoffmann, M., Okamoto, Y. and Woeginger, G.
(2004) The traveling salesman problem with few inner points.
Proc. 10th Int. Computing and Combinatorics Conf.
(COCOON), Series, Lecture Notes in Computer Science, Vol.
3106, pp. 268–277. Oper. Res. Lett., 34, 106–110.
[41] Mohar, B. (1999) A linear time algorithm for embedding graphs
in an arbitrary surface. SIAM J. Discret. Math, 12, 6–26.
[42] Chambers, E., de Verdiere, E., Erickson, J., Lazarus, F.
and Whittlesey, K. (2006) Splitting (complicated) surfaces is
hard. Proc. 22nd Annu. ACM Sympp. Comput. Geom. (SoCG),
pp. 421–429. ACM.
[43] Alber, J. and Fiala, J. (2004) Geometric separation and exact
solutions for the parameterized independent set problem on
disk graphs. J. Algorithms, 52, 134–151.
Page 12 of 13 P. GIANNOPOULOS
THE COMPUTER JOURNAL, 2007
[44] Huffner, F., Niedermeier, R. and Wernicke, S. Techniques for
practical fixed-parameter algorithms. Comput. J. 10.1093/
comjnl/bxm040.
[45] Alber, J., Fernau, H. and Niedermeier, R. (2001) Graph
separators: a parameterized view. Proc. 7th COCOON 2001,
Lecture Notes in Computer Science, Vol. 2108, pp. 318–327.
[46] Marx, D. (2006) Parameterized Complexity of Independence
and Domination on Geometric Graphs. Proc. 2nd Int.
Workshop on Parameterized and Exact Computation
(IWPEC), Lecture Notes in Computer Science, Vol. 4169,
154–165.
[47] Kara, J. and Kratochvıl, J. (2006) Fixed Parameter Tractability
of Independent Set in Segment Intersection Graphs. Proc. 2nd
Int. Workshop on Parameterized and Exact Computation
(IWPEC), Lecture Notes in Computer Science, Vol. 4169,
pp. 166–174.
[48] Dujmovıc, V. et al. (2006) A fixed-parameter approach to
2-layer planarization. Algorithmica, 45, 159–182.
[49] Suderman, M. and Whitesides, S. (2005) Experiments with the
fixed-parameter approach for two-layer planarization. J. Graph
Algorithms Appl., 9, 149–163.
[50] Dujmovıc, V., Fernau, H. and Kaufman, M. (2004) Fixed
parameter algorithms for 1-sided crossing minimization. Proc.
11th Int. Sympp. Graph Drawing (GD ’03), Lecture Notes in
Computer Science, Vol. 2912, pp. 332–344.
[51] Dujmovıc, V. and Whitesides, S. (2004) An efficient fixed
parameter tractable algorithm for 1-sided crossing
minimization. Algorithmica, 40, 15–31.
[52] Dujmovic, V. et al. (2001) On the parameterized complexity of
layered graph drawing. Proc. 5th Annu. European Sympp.
Algorithms (ESA), Lecture Notes in Computer Science, Vol.
2161, pp. 488–499.
[53] Sugiyama, K., Tagawa, S. and Toda, M. (1981) Methods for
visual understanding of hierarchical systems. IEEE Trans.
Sys., Man Cybern., 11, 109–125.
[54] Grohe, M. (2004) Computing crossing numbers in quadratic
time. J. Comput. Syst. Sci., 68, 285–302.
[55] Knauer, C., Schramm, E., Spillner, A. and Wolff, A.
Configurations with Few Crossings in Topological Graphs.
Proc. 16th Annual International Symposium on Algorithms
and Computation (ISAAC), Lecture Notes in Computer
Science, Vol. 3827, pp. 604–613.
[56] Guo, J., Huffner, F. and Niedermeier, R. (2004) A structural
view on parameterizing problems: distance from triviality.
Proc. 1st Int. Workshop Parameterized and Exact
Computation (IWPEC), Lecture Notes in Computer Science,
Vol. 3162, pp. 162–173.
[57] Sharir, M. and Welzl, E. (2006) On the Number of
Crossing-Free Matchings (cycles, and partitions). SODA ’06:
Proc. Seventeenth Annual ACM-SIAM Symp. on Discrete
Algorithm, 860–869, New York, NY, 33 ACM Press.
[58] Spillner, A. (2005) Optimal Convex Partitions of Point Sets with
Few Inner Points. Proc. 17th Canadian Conf. Computational
Geometry (CCCG), Windsor, Ontario, Canada, pp. 39–42.
[59] Kinsey, L. (1993) Topology of Surfaces. Undergraduate Texts
in Mathematics. Springer-Verlag.
[60] Gross, J. and Tucker, T. (1987) Topological Graph Theory. John
Wiley & Sons.
[61] Mohar, B. and Thomassen, C. (2001) Graphs on Surfaces. Johns
Hopkins University Press.
[62] Hatcher, A. (2002) Algebraic Topology. Cambridge University
Press.
[63] Langerman, S. and Morin, P. (2005) Covering things with
things. Discrete Computat. Geom., 33, 717–729.
[64] Downey, R.G. and Fellows, M.R. (1995) Fixed-parameter
tractability and completeness II: on completeness for W[1].
Theor. Comput. Sci., 141, 109–131.
[65] Downey, R.G. and Fellows, M.R. (1995) Fixed-parameter
tractability and completeness I: basic results. SIAM
J. Comput., 24, 873–921.
[66] Marx, D. (2005) Efficient Approximation Schemes for
Geometric Problems? Proc. 13th Annu. European Sympos.
Algorithms (ESA), Vol. 3669, Lecture Notes in Computer
Science, pp. 448–459.
[67] Reif, J. (1986) Complexity of the generalized Mover’s Problem.
In Schwartz, J.T., Sharir, M. and Hopcroft, J.E. (eds), Planning,
Geometry, and Complexity of Robot Motion, pp. 267–281.
Ablex Publishing Corp.
[68] Cesati, M. and Wareham, H. (1995) Parameterized complexity
analysis in robot motion planning. Proc. 25th IEEE Int. Conf.
Systems, Man, and Cybernetics: Vol. 1, pp. 880–885. IEEE Press.
[69] Sharir, M. and Welzl, E. (1996) Rectilinear and Polygonal
p-piercing and p-center Problems. Proc. 12th Annu. ACM
Sympos. Comp. Geometry (SoCG), Philadelphia, PA, pp.
122–132. ACM.
[70] Cabello, S., Giannopoulos, P. and Knauer, C. (2006) On the
parameterized complexity of d-dimensional point set pattern
matching. Proc. 2nd Int. Workshop on Parameterized and
Exact Computation (IWPEC ), Lecture Notes in Computer
Science, Vol. 4169, Philadelphia, PA, USA, pp. 175–183.
[71] Chen, J., Chor, B., Fellows, M., Huang, X., Juedes, D.,
Kanj, I.A. and Xia, G. (2005) Tight lower bounds for
certain parameterized NP-hard problems. Inf. Comput., 201,
216–231.
[72] Indyk, P. (2004) Nearest neighbors in high-dimensional spaces.
In Goodman, J. and O’Rourke, J. (eds), Handbook of Discrete
and Computational Geometry. (2nd edn), pp. 877–892. CRC
Press LLC.
PARAMETERIZED COMPLEXITY OF GEOMETRIC PROBLEMS Page 13 of 13
THE COMPUTER JOURNAL, 2007