# The Author 2008. Published by Oxford University Press on behalf of The British Computer Society. All rights reserved.

For Permissions, please email: journals.permissions@oxfordjournals.org

Advance Access publication on July 28, 2007 doi:10.1093/comjnl/bxm048

Parameterized Complexity and

Approximation Algorithms
DÁNIEL M ARX

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

Institut für Informatik, Humboldt-Universität zu Berlin, Unter den Linden 6, 10099 Berlin, Germany
Corresponding author: dmarx@informatik.hu-berlin.de

Approximation algorithms and parameterized complexity are usually considered to be two separate
ways of dealing with hard algorithmic problems. In this paper, our aim is to investigate how these
two fields can be combined to achieve better algorithms than what any of the two theories could
offer. We discuss the different ways parameterized complexity can be extended to approximation
algorithms, survey results of this type and propose directions for future research.

Keywords: approximation algorithms; parameterized complexity

Received 12 March 2006; revised 11 December 2006

1. INTRODUCTION error is 1%), then in practice it can be as good as an

optimum solution. However, for many approximation algor-
Many of the computational problems that arise in practice are
ithms in the literature, the error guarantee is much higher
optimization problems: the task is to find a solution where the
(50, 100, 1000, 10000%, or even worse). In this case, we
cost, quality, size, profit or some other measure is as large or
cannot argue that this approximation algorithm is almost as
small as possible. The NP-hardness of an optimization
good as an exact algorithm. Nevertheless, such algorithms
problem implies that, unless P ¼ NP, there is no polynomial-
are still important from the theoretical point of view, as they
time algorithm that finds the exact value of the optimum.
allow us to better understand and classify problems and their
Of course, the unfortunate fact that we cannot find the
variants.
optimum in polynomial time does not give us an excuse to
In the literature on approximation algorithms, the goal is
ignore the problem. After all, in practice, some solution is
almost always to find a polynomial-time approximation algor-
required. If we want to design and analyze algorithms for
ithm. However, the question of approximability makes sense
such problems in a mathematically rigorous way, then there
also in the framework of parameterized complexity. By apply-
are several options ahead of us. The field of exact algorithms
ing ideas from both theories, we might be able to tackle pro-
relaxes the requirement that the running time is polynomial,
blems that are intractable to both theories. There can be
and here our aim is to find the algorithm with the fastest
problems that are both inapproximable and fixed-parameter
running time, which is usually exponential in the size of the
intractable, but have parameterized approximation algorithms
input. In parameterized complexity, the running time is ana-
running in fixed-parameter tractable (FPT)-time.
lyzed in finer detail: instead of expressing it as a function of
The aim of this survey paper is to investigate the various
the input size, one or more parameters of the instance are
ways the notion of approximability can appear in parameter-
defined, and we investigate the effect of these parameters on
ized complexity. We review the results on this topic from
the running time. The goal is to design algorithms that work
the literature, and try to interpret these results in a common
efficiently if the parameters of the input instance are small
framework. To motivate further research, we discuss the
(even if the size of the input is large). When designing
different directions that can be pursued. The following four
approximation algorithms, we relax the requirement that the
main issues are identified:
algorithm produces an optimum solution, and our aim is to
devise a polynomial-time algorithm such that the solution it † Approximation with instance parameters: Consider an
produces is not necessarily optimal, but there is some worst- optimization problem, where an integer parameter k is
case bound on the solution quality. associated with every input instance. Is it possible to
The motivation for studying approximation algorithms is find an approximation algorithm with running time
two fold. Firstly, if we have an approximation algorithm f(k) . jxjO(1)? For example, in the PARTIAL VERTEX
whose error guarantee is really good (say, the maximum COVER problem, the task is to cover as many edges as

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

 PARAMETERIZED COMPLEXITY AND APPROXIMATION ALGORITHMS 61

possible with k vertices (where k is part of the input). Is it discussed in Sections 3–6. Conclusions are given in Section 7.
possible to find a 2-approximation in time f(k) . jxjO(1)? Most of the results presented in this paper are taken from
Another example: GRAPH COLORING is fixed-parameter the literature. Some of the cited results are mentioned only
intractable if the parameter is the genus of the graph. informally, while others are presented as theorems. To main-
However, there is a 2-approximation algorithm with tain the flow of the text, only those results are highlighted as
running time f(g) . jxj, if g is the genus of the graph [1]. theorems that are directly related to both parameterization
† Approximation parameterized by cost: The most and approximability.

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

direct application of parameterized complexity to optim-
ization problems is to parameterize by the optimum
value. That is, we try to find an exact algorithm that 2. PRELIMINARIES
solves the problem in f(OPT) . jxjO(1) time, where OPT
2.1 Parameterized complexity
is the value of the optimum solution. The motivation is
that such an algorithm can work efficiently if the We follow [2] for the standard definitions of parameterized
optimum value is very small compared to the size of complexity. Let S be a finite alphabet. A decision problem
the input, which might be a reasonable assumption in is represented by a set Q # S* of strings over S. A parameter-
certain applications. We can investigate the analogous ization of a problem is a polynomial-time computable function
question for approximation algorithms as well. For k : S* ! N. A parameterized decision problem is a pair (Q, k),
example, we might be interested in the question where Q # S* is an arbitrary decision problem and k is a para-
whether there is a 2-approximation algorithm for meterization.1 Intuitively, we can imagine a parameterized
MAXIMUM CLIQUE with running time f(OPT) . jxjO(1). problem as being a decision problem where each input
† Performance functions instead of performance ratios: instance x [ S* has a positive integer k(x) associated with
The usual aim of approximation theory is to find an algor- it. A parameterized problem (Q, k) is FPT if there is an algor-
ithm such that the ratio of the optimum value and the ithm that decides whether x [ Q in time f(k(x)) . jxjc for some
value of the solution provided by the algorithm can be constant c and computable function f. An algorithm with such
bounded. For example, in the case of a minimization running time is called an FPT-time algorithm or simply
problem, the aim is to ensure that the cost of the solution FPT-algorithm. The theory can be extended to problems
is at most c  1 times the optimum, for a constant c as having multiple parameterizations; in this case, we say that a
small as possible. We investigate a more general notion problem (Q, k1, . . . , kp) is FPT with combined parameters
of approximability: instead of the requirement that the k1, . . . , kp if it can be solved in time f(k1(x), . . . , kp(x)) . jxjc.
value of the solution is at most c times the optimum, Many NP-hard problems were investigated in the parame-
we require that it is at most @(OPT) times the optimum terized complexity literature, with the goal of identifying
for some function @. FPT problems. It turns out that several of these problems,
† Quality of approximation as parameter: For certain e.g. MINIMUM VERTEX COVER, LONGEST PATH, DISJOINT
problems, the approximation ratio can be as close to 1 TRIANGLES, etc., are indeed FPT. There is a powerful
as we would like: for every e  0, there is a polynomial- toolbox of techniques for designing FPT-algorithms: kerneli-
time algorithm with approximation ratio 1 þ e . A zation, bounded search trees, color coding, well-quasi-order-
polynomial-time approximation scheme (PTAS) is an ing, just to name some of the more important ones. On the
algorithm that has a parameter e in the input, and pro- other hand, certain problems resisted every attempt at obtain-
duces a (1 þ e )-approximate solution in time jxj f (1/e ). ing FPT-algorithms. Analogously to NP-completeness in
Typically, an approximation scheme has running time classical complexity, the theory of W[1]-hardness can be
4
such as jxj1/e or 21/e . jxj2. To investigate the dependence used to give strong evidence that certain problems are unlikely
of the running time on e , we study the parameterized to be FPT. W[1]-hardness is usually proved by presenting a
version of the problem with 1/e being the parameter. parameterized reduction. For the technical details, the reader
We are especially interested in the question whether is referred to [2, 3].
1/e can be taken out of the exponent of the input size (as
in 21/e . jxj2).
2.2 Optimization problems and approximation
These four issues are not completely separate from each other. For each input instance of an optimization problem, there is a
There are many ways in which they overlap and interact. set of feasible solutions associated to it, and a cost measure is
However, we would like to emphasize here that parameterized
1
approximation is a complex field, and there are many subareas Some authors (e.g., [3]) prefer to define parameterized problems such that
the parameter value is not a function of the instance, but it is a number expli-
to investigate. citly given in the instance. Each approach has its strengths and weaknesses.
Section 2 reviews the basic notions of parameterized com- We prefer to define the parameter as a function, to emphasize that a
plexity and approximation algorithms. The four issues are problem can have many possible parameterizations.

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

62 D. MARX

defined for each feasible solution. The task is to find a feasible schemes were defined that avoid this problem. An efficient
solution where the measure is as good as possible. Following PTAS EPTAS is a PTAS with running time of the form
[4], an NP optimization problem is formally defined as a f(1/e ) . jxjO(1), while a fully PTAS FPTAS runs in time
4-tuple (I, sol, cost, goal), where (1/e )O(1) . jxjO(1).
It is possible to prove negative results on the existence of
† I is the set of instances;
approximation schemes using APX-hardness. APX is the
† for an instance x [ I, sol(x) is the set of feasible solutions
class of optimization problems that have constant-factor
of x. The length of each y [ sol(x) is polynomially

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

approximation algorithms. A PTAS for an APX-hard
bounded in jxj, and it can be decided in polynomial
problem would imply that there is a PTAS for every
time whether y [ sol(x) holds for given x and y;
problem in APX. The celebrated PCP theorem (cf. [4])
† given an instance x and a feasible solution y, cost(x,y) is a
states that this would also imply P ¼ NP. Consequently, for
polynomial-time computable positive integer;
every APX-hard optimization problem there exists a constant
† goal is either min or max.
c0 . 1 such that there is no polynomial-time c0-approximation
The goal is to find a feasible solution y that achieves the best algorithm for the problem, unless P ¼ NP. Often, there is a
objective value, i.e. large gap between this lower bound c0 and the best known
approximation ratio. However,
 for the problem MAX
costðx; yÞ ¼ goalfcostðx; y0 Þ : y0 [ solðxÞg: E3SAT, a tight bound of 87 is proved by Håstad [5].

The cost of the optimum solution for instance x is denoted

by opt(x). If y is a solution for instance x, then the performance 2.4 Decision problems
ratio of y is defined as
In many cases, it is easier to work with decision problems than
8
> costðx; yÞ with optimization problems: most of complexity theory is
>
< ; if goal ¼ min;
optðxÞ developed for decision problems. There is a standard way in
Rðx; yÞ ¼ which an optimization problem can be turned into a more or
> optðxÞ
>
: ; if goal ¼ max : less equivalent decision problem. If X is an optimization
costðx; yÞ
problem, then we define decision problem X as
Thus, R(x, y) is always at least 1; the closer it is to 1, the closer
the solution is to the optimum. For a real number c . 1, we say X
that an algorithm is a c-approximation algorithm, if it always
produces a solution with performance ratio at most c. Input: An instance x of X, an integer k.
Decide: OPT(x)  k.

2.3 Approximation schemes

The problems X and X¼ can be defined analogously. If X
It is a very common situation that after finding the first con- can be solved optimally in polynomial time, then clearly all
stant factor approximation for some problem, improved algor- these decision problems can be solved in polynomial-time as
ithms with better and better approximation ratios are published well. The other direction is not that much clear. If we can
in subsequent papers. For certain problems, an endless series solve X in polynomial time, then OPT(x) can be deter-
of improvements is possible, as there is no ‘best’ approxi- mined in polynomial time using binary search, but this
mation ratio. We say that a problem X admits a PTAS if for does not give us a solution with optimum cost. However,
every e . 0, there is a polynomial-time (1 þ e )-approximation for many problems, if we can determine the optimum,
algorithm for X. More precisely, we want this to hold in a then we can actually find an optimum solution by repeatedly
uniform way: there is one algorithm that can produce an arbi- determining the optimum cost for slightly modified versions
trary good approximation. That is, there is an algorithm A such of the instance, a strategy generally known as polynomial-
that given an instance x of X and an e . 0, A produces a (1 þ time self-reducibility.
e )-approximate solution in time jxj f (1/e ) for some function f. In parameterized complexity, the problems X, X, X¼ are
Clearly, such an algorithm runs in polynomial time for every studied with the parameter k. If X is a minimization (respect-
fixed value of e . Approximation schemes are very abundant ively, maximization) problem, then the standard parameteri-
for geometric problems, but the literature contains many zation of X is the problem X (respectively, X)
examples for other types of problems as well, e.g. for schedul- parameterized by k. If the standard parameterization is FPT,
ing, packing or pattern matching problems. then this means that we have an efficient algorithm for deter-
If e is small then the exponent of the polynomial jxj f (1/e ) can mining the optimum for those instances where the optimum is
be very large. Two restricted classes of approximation small.

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

 PARAMETERIZED COMPLEXITY AND APPROXIMATION ALGORITHMS 63

3. APPROXIMATION WITH INSTANCE The traveling salesperson problem (TSP) is one of the most
PARAMETERS studied optimization problems. The input consists of n cities
and a matrix describing the distances between the cities (the
The first way of studying parameterized approximation algor-
distances can be arbitrary nonnegative integers). The task is
ithms that we discuss here is to define a parameterization of
to start from one of the cities, visit every city in some order
the optimization problem, and instead of looking for a
and then return to the starting city. We have to find an ordering
polynomial-time algorithm, our goal is to find an FPT-time
of the cities such that the length of this tour is as small as
algorithm that produces an approximate solution. That is, we

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

possible. The problem is NP-hard, in fact, there is no
define a parameterization k that assigns a positive integer to
c-approximation algorithm for any c  1 (unless P ¼ NP).
each instance x [ I, and the approximation algorithm should
The Metric TSP is a restricted version of TSP where we
run in time f(k(x)) . jxjO(1). In this case, we say that the algor-
assume that the distance matrix satisfies the triangle inequality
ithm is an FPT-approximation algorithm with parameteri-
d(x, y)  d(x, z) þ d(z, y) for any three cities x, y, z. This
zation k. If the performance ratio of the algorithm is c, then
assumption obviously holds in applications where we know
we say that it is an FPT c-approximation algorithm. Hope-
that the direct route is never longer than a route via multiple
fully, such an algorithm can achieve a better approximation
cities. Metric TSP can be 1.5-approximated using Christo-
ratio than traditional polynomial-time approximation algor-
fides’ algorithm [8], but has no PTAS, unless P ¼ NP [9].
ithms, as it has more time at its disposal.
METRIC TSP WITH DEADLINES (DDLTSP) is investigated in
The following two results are examples of this type of para-
[6]. Here we have the further restriction that some subset D of
meterized approximation algorithms (definitions will be given
the vertices have deadlines assigned to them, and we have to
later).
visit each such vertex before its deadline. That is, each v [
† METRIC TSP WITH DEADLINES (DDLTSP) can be D has a deadline d(v), and the length of the tour from the start-
2.5-approximated in time O(jxj3 þ k!k), where k is the ing city to v must not exceed d(v). The task is to find the short-
number of deadline vertices [6]. est tour that satisfies all the deadlines. The problem has no
† VERTEX COLORING can be 2-approximated in f(g) . jxj polynomial-time constant factor approximation algorithm,
time for graphs with genus g [1]. unless P ¼ NP. Furthermore, it is not FPT if the parameter
is the number jDj of deadline vertices; in fact, it is NP-hard
Broadly speaking, there are two large categories of parameter-
even if there is only one deadline vertex. Thus, neither
izations: the parameter is either some obvious measure of the
approximation nor parameterized complexity alone can
problem instance (e.g. the number of deadline vertices in
tackle this problem. However, [6] presents a O(jxj3 þ k!k)
DDLTSP) or it is some structural property of the input struc-
time algorithm that produces a 2.5-approximate solution, if k
ture that describes (in some well-defined sense) how compli-
is the number of deadline vertices. Therefore, the combination
cated the input is (e.g. tree width, genus, etc.) In both cases,
of approximation and parameterized complexity results in an
it can be interesting to find approximation algorithms that
efficient approximation algorithm for small values of k.
are efficient for small parameter values. Sections 3.1 and 3.2
discuss results of these two types, respectively.
THEOREM 3.1 [6]. METRIC TSP WITH DEADLINES parameter-
Parameterized approximation schemes can be defined in a
ized by the number of deadline vertices has an FPT
similar manner. An FPT-approximation scheme (FPT-AS)
2.5-approximation algorithm.
with parameterization k is an algorithm whose input is an
instance x [ I and an e . 0, and it produces a (1 þ e )-
The VERTEX COVER problem (cover all the edges of a graph
approximate solution in time f(e ,k(x)) . jxjO(1) for some com-
with as few vertices as possible) is well studied both in the
putable function f. For example, Har-Peled and Mazumdar
parameterized complexity and approximation algorithms lit-
[7] present an f(e ,k,d) . jxj time approximation scheme for
erature. VERTEX COVER is FPT (see, e.g. [10, 11]) and has a
the MINIMUM k-MEDIAN problem in d dimensions, which
simple 2-approximation algorithm. However, the problem
means that the problem has an FPT-AS with combined par-
does not have a PTAS (unless P ¼ NP); in fact, some recent
ameters k and d. This can be considered as a generalization
conjectures imply that the factor 2 is best possible [12].
of EPTAS: the constant factor depends not only on e , but
PARTIAL VERTEX COVER is in some sense the dual of
also on the parameters of the instance. Sections 3.1 and 3.2
VERTEX COVER: given a graph G with an integer k, the task
review this type of approximation results as well.
is to cover as many edges as possible with k vertices. It can
be shown that this problem is W[1]-hard with parameter k
[13]. The straightforward greedy algorithm (repeatedly
3.1. Measure parameters
select a vertex that covers as many uncovered edges as poss-
In this section, we overview parameterized approximation ible) gives a 1.582-approximation [14], but the problem does
results where the parameter is some obvious measure of the not admit a PTAS [15]. Here we show that the problem
input instance. admits an FPT-AS.

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

64 D. MARX

THEOREM 3.2. PARTIAL VERTEX COVER admits an FPT-AS PTAS. The computational geometry community has been
with parameter k, the number of vertices in the solution. doing parameterized approximation analysis for several
years: the quest for obtaining approximation algorithms
Proof. We present an f(e ,k) . jxjO(1) time algorithm that pro- where certain constants do not appear in the exponent of the
duces a (1 þ e )-approximate solution for the problem. Let input size is evident in the literature.
 
D :¼ 2 2k e and let v1, . . . , vn be the vertices of the graph The MINIMUM k-CENTER problem arises as a fundamental
ordered by non-increasing degree, i.e. d(vi)  d(vj) for i , j. problem in many applications such as facility location, cluster-

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

We consider two cases. ing and information retrieval. Given a set of points in the plane
and an integer k, the task is to select k center points such that
Case 1: d(v1)  D. In this case the algorithm outputs the set each input point is close to some selected point. More pre-
S ¼ fv , . . . , vkg. These k vertices cover at least cisely, we have to minimize the maximum distance of a
Pk 1 k  k  point to the closest selected point.
i¼1 dðvi Þ  2 edges: there are at most 2 edges between
k 
the vertices of S, thus at most 2 edges are counted twice
when the degrees are summed. MINIMUM k-CENTER
P The value of the optimum sol- Input: A set S # R  R of points, an integer k.
ution cannot be larger than ki¼1 d(vi), hence the value of the
constructed solution S is at least Find: A subset C # S of size k.
Goal: Minimize maxs[S minc[C d(s,c).
   
Pk k k
i¼1 dðvi Þ  2 2 e 1 There are several ways of interpreting distance in the
Pk 1 ¼1 
dðv i Þ D 2 1 þ e plane. Here we discuss the two most common cases,
i¼1
where the distance means ffi L2 metric (d((x1, y1), (x2, y2)) ¼
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
jx1 2 x2 j2 þ jy1 2 y2 j2 Þ or L1 metric (d((x1, y1), (x2,
times the optimum, that is, the performance ratio is at most
y2)) ¼ maxfjx1 2 x2j,jy1 2 y2jg). In both cases, the problem
1 þ e.
can be 2-approximated in polynomial time, but there is some
constant a . 2 such that it is NP-hard to find an
Case 2: d(v1)  D. In this case, the optimum value is at most
a-approximation [17]. If the number of centers is a fixed con-
kD, and we are able to determine it exactly in FPT-time. For
stant k, then the problem can be solved exactly in time jxjO(k)
each 1  ‘  kD, the algorithm checks whether it is possible
by brute force. This cannot be improved to f(k) . jxjO(1) time:
to cover at least ‘ edges with k vertices. We use the method
parameterized by the number k of centers, the problem is
of color coding: random colors between 1 and ‘ are assigned
W[1]-hard in the L1 metric [18]. However, there are efficient
to the edges, and we try to find a solution where the ‘ edges
approximations for every fixed value of k: [19] shows that for
covered by the vertices have distinct colors. As shown
every fixed k there is a linear-time approximation scheme, i.e.
below, the existence of such a solution can be checked in
a (1 þ e )-approximate solution can be found in f(e , k) . n time.
FPT-time for a given coloring. If it is not possible to cover ‘
Using the terminology of parameterized approximation, this
edges, then obviously the algorithm does not find a solution
can be stated as follows.
for any random coloring. On the other hand, if there is a sol-
ution, then the random coloring assigns distinct colors to the THEOREM 3.3. [19]. MINIMUM k-CENTER admits an FPT-AS
‘ edges with probability at least ‘2‘. Therefore, we have to with parameter k.
repeat the algorithm on average at most ‘‘ times to find a sol-
MINIMUM k-MEDIAN is the variant of the problem where
ution. Note that ‘‘ can be bounded by a function of e and k
instead of minimizing the maximum distance to the closest
only. The algorithm can be derandomized using standard tech-
center point, the optimization goal is to minimize the sum of
niques, (see [2, Section 13.3; 3, Section 8.3; 16]).
these distances. Unlike MINIMUM k-CENTER, this problem
For a particular coloring, we proceed as follows. We con-
admits an EPTAS in the L2 metric (even if k is part of the
sider every possible partition P ¼ fP1, . . . , Pkg of the color
input) [7].
set f1, . . . , ‘g into k classes; there are at most k ‘ such
Most geometric problems can easily be generalized to
partitions. For a given partition P, we try to find vertices
higher dimensions. If the dimension of the input can be arbi-
u1, . . . , uk such that for each color c [ Pi, vertex ui covers
trarily large, then usually the problem becomes as hard as on
at least one edge with color c. This can easily be done in poly-
any general metric, since we lose the nice geometric proper-
nomial time. A
ties that make approximation easy (cf. [20]). However, in
One of the most fruitful areas for developing approximation many cases the geometric approximation schemes can be
algorithms is the field of geometric problems. The geometric generalized to any fixed dimension d. The approximation
nature of the problem allows us to use all sorts of approxi- schemes for MINIMUM k-CENTER [19] and for MINIMUM
mation techniques. Very often, the problem even admits a k-MEDIAN [7] work for an arbitrarily fixed dimension d,

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

 PARAMETERIZED COMPLEXITY AND APPROXIMATION ALGORITHMS 65

and the running time is f(e ,k, d) . jxjO(1) and f(e ,d) . jxjO(1), † Distance from a class F. Let F be an arbitrary class of
respectively. graphs (such as planar, bipartite, interval, etc.) The class
F þ kv (respectively, F þ ke) contains those graphs that
THEOREM 3.4 [7, 19]. MINIMUM k-MEDIAN admits an FPT-AS
can be constructed from some G [ F with the addition of
with parameter d. MINIMUM k-CENTER admits an FPT-AS with
k new vertices2 (respectively, edges). The classes F 2 kv
combined parameters d and k.
and F 2 ke are similarly defined. For any problem and
Dimension is a very natural parameter for the study of geo- any class F, we can assume that the input is in, say,

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

metric problems. In the examples above, FPT-algorithms are F þ kv for some k, and define this k to be the parameter.
possible if the dimension is taken as parameter. Conceivably, This line of research was initiated by Cai [24], and later
there can be problems where it is possible to show that the pursued in, e.g. [25, 26].
exponent of the input size has to increase as the dimension
increases. There is very little work done in the literature on The first problem that we investigate here is VERTEX COLOR-
proving hardness results for problems parameterized by the ING: given a graph G, find a coloring of the vertices with
dimension of the space. The only such result that we are minimal number of colors such that adjacent vertices receive
aware of is on the SUBSET CONGRUENCE problem. Given different colors. The minimum number of colors that is
two sets of points A, B # Rd, the task is to decide if there is required to color the vertices of G is called the chromatic
a distance-preserving transformation that makes A equal to a number of G and is denoted by x(G). Chromatic number is
subset of B. The problem is polynomial-time solvable for not approximable within jVj12e for any e . 0, unless ZPP ¼
every fixed d, but W[1]-hard parameterized by the dimension NP [27]. The problem remains hard for planar graphs: it is
d [21]. NP-complete to decide whether a planar graph is 3-colorable
[28]. However, by the celebrated Four Color Theorem [29],
four colors are sufficient for every planar graph, and there is
3.2. Structural parameters a polynomial-time algorithm that actually finds this coloring
[30].
4 As it is easy to check whether a graph is 2-colorable, a
In the previous section, we discussed parameters that measure
3 -approximate coloring can be found in polynomial time
the size of some part of the input or the solution. These par- for planar graphs.
ameters were ‘obvious.’ Parameterization can be more subtle Is there a constant factor approximation algorithm for
than that: it can measure arbitrary properties of the input struc- graphs that are ‘almost’ planar graphs? We consider two
tures. In many cases, the parameter measures how complicated classes of graphs that are close to being planar: bounded
the input is, in some precise, well-defined sense. The motiv- genus graphs and planarþkv graphs. For bounded genus
ation is to obtain algorithms that are efficient for less graphs, the following result is implicit in [1]:
complex instances, where less complex means that the par-
ameter value is small. Such an algorithm can be useful in THEOREM 3.5 [1]. VERTEX COLORING has an FPT 2-
applications where it is a reasonable assumption that the approximation algorithm if the parameter is the genus of the
input is not completely unstructured, and the value of this graph.
particular parameter is usually small.
Below we list some parameters that describe complexity in 7For planarþkv graphs, an easy brute force algorithm gives a
graphs. The list is not exhaustive; depending on the problem or 3 -approximation:
application, many such parameters can be defined.
7THEOREM 3.6. VERTEX COLORING has an FPT
† Maximum degree. The maximum degree of a graph is 3 -approximation algorithm for planarþkv graphs.
the maximum number of neighbors a vertex can have.
† Diameter. The diameter of a graph is the maximum Proof. Let X ¼ fv1, . . . , vkg be k vertices of G such that G\X is
distance between any two vertices. planar (here we gloss over the question how this set is found or
† Tree width. The tree width is a parameter that describes whether it is given in the input together with the planarþkv
how ‘tree-like’ the graph is. We omit the technical graph, cf. [31]). We can determine x(G[X]) by trying all
definition, see, e.g. [22] for more details. In many the k
cases, problems on graphs with small tree width can be 4 k possible colorings of G[X]. As discussed above, a
3 -approximate coloring of the planar graph G\X can be
handled with dynamic programming techniques. found in polynomial time. The coloring on G[X] and the color-
† Genus. The genus of a graph describes how close the ing on G\X can be combined
graph is to being a planar graph. A graph has genus 0  to obtain a coloring of the whole
graph with c ¼ x(G[X])þ 43 . x(G\X) colors. It is easy to
if and only if it is planar (i.e. can be drawn on the check that c  73 x(G); the worst case occurs when x(G) ¼
sphere without crossing edges). A graph has genus at
most k if it can be drawn on the sphere with k 2
More precisely, we add k new vertices and connect them with each other
‘handles’ attached to the sphere (cf. [23, Appendix B]). and with the old vertices arbitrarily.

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

66 D. MARX

x(G[X]) ¼ x(G\X) ¼ 3, but the algorithm produces a coloring of the standard, well-studied problems in the parameterized
with 3 þ 4 ¼ 7 colors. A complexity literature are standard parameterizations of
certain optimization problems (e.g. MINIMUM VERTEX
Note that Theorems 3.5 and 3.6 are incomparable: bounded
COVER, MAXIMUM CLIQUE, MINIMUM DOMINATING SET,
genus graphs might not be planarþkv for any k (e.g. large tor-
LONGEST PATH, etc.) By studying the fixed-parameter tract-
oidal grids) and planarþ1v graphs can have unbounded genus
ability of these problems, we are investigating the possibility
(e.g. a grid with an additional vertex connected to every
of having efficient exact algorithms for these problems in
vertex).

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

the case when the optimum is small. When a W[1]-hardness
The MAXIMUM INDEPENDENT SET problem is NP-hard for
result shows that such an exact algorithm is unlikely to
planar graphs [32], but has a linear-time approximation
exist, then it is natural to study whether it is possible to
scheme [33, 34]. It follows from [35] and [36] that this can
approximate the optimum, if it is small.
be generalized to bounded-genus graphs.
Recently, at least three papers tried to extend parameterized
THEOREM 3.7 [35, 36]. Maximum independent set admits a complexity into this direction [44 – 46]. In Section 3, the defi-
linear-time FPT-AS if the parameter is the genus of the graph. nition of FPT-approximability (when the parameter is some
property of the optimization instance) was a straightforward
For planarþkv graphs, a (1 þ e )-approximation of
generalization of polynomial-time approximability. The defi-
maximum independent set can be found by trying all 2k possi-
nition becomes technically more delicate if we want to para-
bilities on the k-extra vertices, and then by finding a (1 þ e )-
meterize by cost. The first complication is that we have to
approximation on the remaining graph.
decide whether we want to parameterize by the optimum
THEOREM 3.8. MAXIMUM INDEPENDENT SET admits a linear- cost (which is somewhat counterintuitive, since presumably
time FPT-AS on planarþkv graphs. the optimum cost is hard to determine), or we assume that
the input contains a parameter k (the cost that should be
In VERTEX COLORING, the goal is to minimize the number
reached), as in the standard parameterization of the problem.
of different colors used. In other words, if we identify the
Here we follow the definition proposed by Chen et al. [45].
set of colors with the set of positive integers, then we try to
minimize the maximum of the colors assigned. In MINIMUM DEFINITION 4.1. Let X ¼ (I, sol, cost, goal) be an optimiz-
SUM COLORING, the optimization goal is to minimize the ation problem. A standard FPT-approximation algorithm
sum of the colors (positive integers) on the vertices. Besides with performance ratio c for X is an algorithm that, given
its combinatorial interest, the problem is motivated by appli- an input (x, k)[ S*  N satisfying
cations in scheduling and VLSI design [37, 38]. The 
MINIMUM SUM EDGE COLORING problem is defined analo- OPTðxÞ  k; if goal ¼ min;
ð1Þ
gously. The MINIMUM SUM COLORING is linear-time solvable OPTðxÞ  k; if goal ¼ max
on bounded tree width graphs [39], but the MINIMUM SUM
EDGE COLORING is one of the few problems that are
polynomial-time solvable on trees [40, 41], but NP-hard
already on graphs with tree width 2 [42]. However, computes a y [ sol(x) in time f(k) . jxjO(1) such that
MINIMUM SUM EDGE COLORING has a linear-time PTAS for 
bounded tree width graphs (in fact, this is true even for the costðx; yÞ  k  c; if goal ¼ min;
ð2Þ
much more general multicoloring version). costðx; yÞ  k=c; if goal ¼ max :

THEOREM 3.9 [43]. MINIMUM SUM EDGE COLORING admits

For inputs not satisfying condition (1), the output can be
an FPT-AS if the parameter is the tree width.
arbitrary.
Note that without parameterization we cannot obtain an
The word ‘standard’ signifies that this is the approximation
approximation scheme for MINIMUM SUM EDGE COLORING,
version of the standard parameterization, and distinguishes it
since the problem is APX-hard [42].
from the case when the parameter is unrelated to the cost (as
in Section 3).
Unfortunately, we do not have a good example for this form
4. PARAMETERIZATION BY COST
of approximability, an example where a problem is not FPT,
In Section 3, our aim was to develop approximation algor- but has a standard FPT-approximation algorithm with some
ithms that work efficiently if some parameter of the input is performance ratio c . 1. Nevertheless, there are some
small. Perhaps the most obvious parameter of an optimization examples of FPT-approximation algorithms appearing in the
problem instance is the optimum cost. What we would like to literature. Tree width is FPT, in fact for every k, a tree
have is an algorithm that efficiently finds an approximation decomposition of width k can be computed in linear time if
of the optimum, if this optimum is small. Notice that many exists [47]. This algorithm is quite complicated and not

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

 PARAMETERIZED COMPLEXITY AND APPROXIMATION ALGORITHMS 67

practical. However, there are much simpler FPT-algorithms ratio function @(k), (2) has to be replaced with
that produce tree decompositions having width at most a con- 
stant factor larger than the optimum [48 – 57]. In the case of the costðx; yÞ  k  @ðkÞ; if goal ¼ min;
rank width of graphs and branch width of matroids represented costðx; yÞ  k=@ðkÞ; if goal ¼ max :
over a fixed-finite field, we have the curious situation that it
can be decided in FPT-time whether the width is at most k If our goal is only to obtain an approximation algorithm
[52, 53], but it is not known whether a decomposition of with some performance function (no matter how bad it is),

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

width at most k (if it exists) can be constructed in FPT-time. then in the case of maximization problems, FPT time does
However, standard FPT-approximation algorithms exist for not give us more power than polynomial time. This surprising
these problems [53– 55]. result was observed by Grohe and Grüber [56].
One might be tempted to define the notion of standard
FPT-AS to capture the situation where a problem has a standard THEOREM 5.1. If a maximization problem X has a standard
FPT-approximation algorithm for every c . 1. More precisely, FPT-approximation algorithm with performance ratio func-
we would like to have an analog of EPTAS, since approxi- tion @(k), then there is a polynomial-time @ 0 (k)-approximation
mation schemes where the exponent of input size increases algorithm for X, for some function @ 0 (k).
as e decreases is not very interesting from the parameterized
complexity point of view. Thus, the natural definition for a Proof. Assume that X has a standard FPT-approximation
standard FPT-approximation scheme would be that for every algorithm with running time f(k) . jxjO(1) and performance
e . 0, there is a standard FPT-approximation algorithm with ratio function @(k). There is a computable nondecreasing func-
performance ratio 1 þ e , and the running time is f(k,e ) . tion b(n) such that every instance of X with size at most n has
jxjO(1). However, this notion is not very interesting to study, optimum at most b(n). Denote by c(n) the largest i such that
since it implies fixed-parameter tractability; therefore, it does i  n, f(i)  n and f(i) can be computed in time O(n);
not extend the class of problems that we can call tractable. clearly, c(n) is unbounded, nondecreasing and computable in
The proof is essentially the same as the proof of Proposition 6.1. time n O(1).
Given an instance x, the polynomial-time approximation
PROPOSITION 4.1. Assume that there is an algorithm that has an algorithm proceeds as follows. For every k ¼ 1,2, . . . ,
input parameter e , and for every e . 0 it produces a (1 þ e )- max(c(jxj),1), we run the assumed FPT-approximation algor-
approximate solution for optimization problem X in time f(k,e ) . ithm on x, and select the best solution returned by the different
jxjO(1). Then the standard parameterization of X is in FPT. runs (since we try k ¼ 1 and OPT(x)  1 holds, some solution
is always found). From k  max(c(jxj),1) it follows that the
f(k) . jxjO(1) time FPT-algorithm runs in polynomial time,
hence the running time of the whole procedure is polynomial
5. NON-CONSTANT PERFORMANCE FUNCTIONS in jxj.
When designing approximation algorithms, the usual aim is to If c(jxj)  OPT(x), then the FPT-approximation algorithm is
ensure that the cost of the solution differs from the optimum by invoked with k ¼ OPT(x), which means that a solution of cost at
at most a constant factor. This means that the algorithm has to least OPT(x)/@(OPT(x)) is produced. Assume therefore that
perform equally well regardless of whether the optimum is c(jxj) , OPT(x). Let d(m) be the largest n such that b(n) , m.
‘small’ or ‘large.’ The analysis can be made more precise if This means that jxj. d(OPT(x)) for every instance x, hence
the performance of the algorithm is bounded by a function c(jxj) c(d(OPT(x)). Thus, when the FPT-approximation algor-
of the optimum cost, rather than by an absolute constant. ithm is invoked with k ¼ c(d(OPT(x))) , OPT(x), then it
This is especially natural in the case of standard produces a solution with value at least r(OPT(x)) :¼
FPT-approximation (Section 4): if the goal is to have an algor- c(d(OPT(x)))/@(c(d(OPT(x)))). It is easy to see that r(OPT(x))
ithm that is efficient for small cost values, then it makes sense is an unbounded computable function of OPT(x), hence the
to require good performance only if the optimum is small. polynomial-time algorithm is an approximation algorithm with
Let @: N ! N be a nondecreasing function. We say that an performance ratio function @0 (k) ¼ k/r(k). A
approximation algorithm for optimization problem X has We remark that the analog of Theorem 5.1 does not seem to
performance ratio function @(k) (or it is a @(k)-approximation hold for minimization problems.
algorithm) if it produces a solution with performance ratio at
most @(OPT(x)) for every instance x. For example, if the solution
has cost OPT(x) log OPT(x), then it is an log-approximation
5.1. Examples (polynomial-time)
algorithm. Constant-factor approximability is the special case
where @(k) is a constant function. The definition of parameter- Expressing the performance ratio of an algorithm as a function
ized FPT-approximation (Section 3) can be extended similarly. of the optimum can be interesting even in the case of
To define standard FPT-approximability with performance polynomial-time approximation algorithms, independently of

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

68 D. MARX

parameterized complexity. We cite here three results of this the layout. The cutwidth of a layout is max1i,jVjjfuv [ E :
type. Determining tree width is NP-hard [57], and no constant s(u) i and s(v) .igj, the maximum number of edges that
factor polynomial-time approximation algorithm is known cross the cut formed by two neighboring vertices in the
(it is an important open question whether such an approxi- layout. The bandwidth (respectively cutwidth) of a graph is
mation is possible). However, the following approximation the minimum bandwidth (respectively cutwidth) over all poss-
result was obtained recently by Feige et al. [58]. ible linear layouts of the graph. Bandwidth is W[1]-hard [66],
while cutwidth is FPT [67].
THEOREM 5.2 [58]. Given a graph with tree width k, a tree
pffiffiffiffiffiffiffiffiffiffi

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

Subdividing an edge with a new degree two vertex might
decomposition of width at most Oðk log kÞ can be con-
decrease the bandwidth of the graph. Topological bandwidth
structed in polynomial time.
pffiffiffiffiffiffiffiffiffiffi is the smallest bandwidth that can be achieved by repeatedly
That is, tree width can be Oð log kÞ-approximated in subdividing edges. Topological bandwidth is known to be
polynomial time. (The previous best result was a O(log k)- W[1]-hard [68]. Fellows [69] observed that if tbw(G) (respect-
approximation [51, 59].) ively cw(G)) is the topological bandwidth (respectively
Whether MINIMUM DIRECTED FEEDBACK VERTEX SET is in cutwidth) of graph G, then
FPT or not is one of the most important open question of para-
meterized complexity. In this problem, the task is to find a tbwðGÞ  cwðGÞ þ 1 ð3Þ
minimum set of vertices whose deletion makes the given
directed graph acyclic. The undirected version of the
problem is known to be in FPT (cf. [3, 60]), but apparently and
none of the techniques go through for the directed case. It
seems that a much deeper understanding of the structure of
cwðGÞ , tbwðGÞ2 ð4Þ
directed graphs is required before this problem can be
settled. However, it is observed in [45] that the results of
[61, 62] imply an approximation algorithm with bounded per- holds. This means that if a graph has topological bandwidth at
formance ratio function. most k, then one can obtain a layout with cutwidth less than k 2
using the algorithm of [67]. Furthermore, using (3) (whose
THEOREM 5.3 [45]. MINIMUM DIRECTED FEEDBACK VERTEX
proof is algorithmic), this layout can be turned into a layout
SET has a polynomial-time approximation algorithm with per-
for a subdivision of G with bandwidth at most k 2.
formance ratio function O(log k . log log k).
THEOREM 5.5 [69]. TOPOLOGICAL BANDWIDTH has a standard
In the MAXIMUM DISJOINT CYCLES problem, the task is to
FPT-approximation algorithm with performance ratio function k.
find k pairwise vertex disjoint cycles in a given graph. This
problem is FPT (cf. [3 Section 8.1]). On the other hand, the The graph parameter clique width was introduced by Cour-
directed version of the problem is W[1]-hard [45, 63], but celle et al. [70]. Similarly to tree width, it measures the com-
polynomial-time solvable for every fixed value of k [64]. On plexity of a graph with respect to certain composition
the basis of the theoretical results proved in [64], Grohe and operators. A k-expression describes a way of constructing
Grüber [56] presented an FPT-approximation algorithm. By the graph using these operations; the clique width of a graph
Theorem 5.1, the algorithm can be made to work in poly- is the smallest k such that it has a k-expression. Fellows
nomial time. et al. [71] have shown that determining clique width is
NP-hard, in fact, it cannot be approximated with an absolute
THEOREM 5.4 [56]. There is a polynomial-time approxi-
error guarantee, unless P ¼ NP. Oum and Seymour [54] use
mation algorithm for MAXIMUM DISJOINT DIRECTED CYCLES
the approximability of rank width to obtain an approximation
with performance ratio function @(k), for some computable
algorithm for clique width. They show that if cw(G) is the
function @(k).
clique width and rw(G) is the rank width of a graph G, then
The function @(k) is not given explicitely in [56], but k/@(k)
is very slowly growing (even in the FPT-time version of the
rwðGÞ  cwðGÞ  2rwðGÞþ1  1
algorithm, before applying Theorem 5.1.).

holds. The proof is algorithmic: given a rank decomposition of

5.2. Examples (FPT-time)
width k, a (2kþ1 2 1)-expression can be obtained in poly-
Graph layout problems arise in many application domains nomial time. There is an FPT-approximation algorithm that
such as scheduling, VLSI design and archeology [65]. A computes a rank decomposition of width 3rw(G) þ1 [55].
linear layout of a graph G(V, E) is a one-to-one mapping s Therefore, clique width can be approximated by first comput-
between V and f1, . . . , jVjg. The bandwidth of a layout is ing an approximate rank decomposition and then turning it
maxuv[Ejs(u) 2 s(v)j, the maximum ‘length’ of an edge in into a k-expression.

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

 PARAMETERIZED COMPLEXITY AND APPROXIMATION ALGORITHMS 69

THEOREM 5.6 [55]. CLIQUE WIDTH has a standard has cost k, which makes approximation equivalent to finding
FPT-approximation algorithm with performance ratio func- an optimum solution. The inapproximability proofs in these
tion (23kþ2 2 1)/k. examples tell us more about the hardness of finding exact
solutions than about the hardness of approximation. It would
The significance of Theorem 5.6 comes from the fact that
be much more interesting (and possibly, more difficult) to
certain optimization problems are linear-time solvable for
have analogous inapproximability results for the monotone
every k, if a k-expression of the graph is given in the input.
problems MAXIMUM CLIQUE and MINIMUM DOMINATING
That is, these problems are FPT parameterized by clique

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

SET, for example.
width, but only if the corresponding k-expression is given in
the input. Theorem 5.6 can be used to remove the requirement
that the k-expression is given in the input: if we know that the
graph has a k-expression, then an f(k)-expression can be 6. QUALITY OF APPROXIMATION
obtained in FPT-time, where f(k) ¼ 2 3kþ2 2 1. Then, instead AS PARAMETER
of using the algorithm for k-expressions, we can use the algor- A PTAS can produce solutions with approximation ratio arbi-
ithm for f(k)-expressions: the running time might be much trary close to 1. However, we have to pay a price for that: the
larger, but it is still linear with a constant depending only on k. running time can ridiculously be large if e is small. Table 1
Therefore, obtaining the approximation and then solving the (reproduced from [72]) presents the running time of some
problem using the f(k)-expression can be done in FPT-time. approximation schemes for e ¼ 0.2. As we can see, running
times of O(jxj1,000,000) or worse is not uncommon if we
require that the maximum error is 20%. Obviously, such algor-
5.3. Negative results ithms are not useful in practice. Nevertheless, these results are
For certain problems, it is possible to show that there is no important from the theoretical point of view, as they show that
standard FPT-approximation algorithm for any performance there is no lower bound on the approximation ratio that can be
ratio function. Given a graph G(V, E), an independent domi- achieved in polynomial time.
nating set is an independent set S # V of vertices such that The notion of EPTAS tries to formalize a restricted class of
for every v [ V, either v [ S, or v is a neighbor of a approximation schemes, where the algorithm might have a
member of S. The MINIMUM INDEPENDENT DOMINATING SET chance of being useful. If the running time of the algorithm
problem is the corresponding optimization problem, where is of the form c 1/e . n, and c is not too large, then the algorithm
the goal is to minimize the size of the set S. Downey et al. can be efficient for, say, e ¼ 0.2. In many cases, the first
[46] prove that this problem is completely inapproximable. approximation scheme obtained for a problem was not an
EPTAS, but later it was improved to an EPTAS. For
THEOREM 5.7 [46]. If MINIMUM INDEPENDENT DOMINATING example, Arora [73] presented an jxjO(1/e ) time PTAS for
SET has a standard FPT-approximation algorithm with per- Euclidean TSP, which is not an EPTAS. However, in the
formance ratio function @(k) for some computable function journal version of the paper [74], the running-time of the
@(k), then W[2] ¼ FPT. 2
algorithm is improved to jxj . logO(1/e ) jxj ¼ 2O(1/e ) . jxj2,
Chen et al. [45] presents inapproximability results for satis- hence the problem admits an EPTAS. Arora et al. [75] pre-
fiability problems. In the MIN-WSAT(CIRC) problem, a sented an jxjO(1/e ) time approximation scheme for the EUCLI-
DEAN k-MEDIAN problem. Later this was improved to an
Boolean circuit is given and the task is to find a satisfying
assignment of minimum weight (where the weight of an EPTAS with running time 2O(1/e log 1/e ) . n [7, 75]. In certain
assignment is the number of true variables). cases, the effort to obtain an EPTAS was only partially suc-
cessful. For example, Hunt et al. [76] show that MAXIMUM
THEOREM 5.8 [45]. If MIN-WSAT(CIRC) has a standard
FPT-approximation algorithm with performance ratio func-
tion @(k) for some computable function @(k), then W[P] ¼
FPT. TABLE 1. The running time of some PTASs with 20% error.
Similar results hold for other variants of the satisfiability Running time
problem and for the parameterized halting problem [45]. Problem Reference for 20% error
All the inapproximability results presented above are some-
EUCLIDEAN TSP Arora [73] O(jxj15,000)
what unsatisfying in the sense that the problems considered are
MULTIPLE KNAPSACK Chekuri and Khanna [111] O(jxj9,375,000)
not monotone. Monotone means that (in case of a minimiz-
MAXIMUM SUBFOREST Shamir and Tsur [112] O(jxj958,267,391)
ation problem) if we extend a feasible solution with additional
MAXIMUM Erlebach et al. [113] O(jxj523,804)
vertices/true variables, then it remains feasible. Clearly, this
INDEPENDENT SET for
does not necessarily hold in these examples. Therefore, it
disk graphs
can happen that the optimum is k, and every feasible solution

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

70 D. MARX

INDEPENDENT SET for unit disk graphs admits an jxjO(1/e ) time TABLE 2. Problems where a W[1]-hardness result rules out the
PTAS. They are unable to present an EPTAS for the problem possibility of an EPTAS.
in general, but for the special case of l-precision unit disk
Problem PTAS result W[1]-hardness
graphs (where the centers of the disk are not closer than l
2
from each other) they give a 2O(1/(e l) ) . n time EPTAS. MAXIMUM INDEPENDENT SET for jxjO(1/e ) [76] [18]
If a problem resists every attempt to obtain an EPTAS, then unit disks
we should consider looking for some negative evidence MAXIMUM INDEPENDENT SET for jxjO(1/e )[76] [18]

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

showing that no EPTAS is possible, i.e. that it is not possible unit squares
to get 1/e out of the exponent of the input size (modulo some MAXIMUM DOMINATING SET for jxjO(1/e ) [76] [84]
complexity-theoretic assumption). It is not very surprising that unit disks
parameterized complexity can provide such evidence, since MAXIMUM DOMINATING SET for jxjO(1/e ) [76] [84]
getting out certain parameters from the exponent is the unit squares
2
central issue of the theory. The basic tool that can be used to COVERING POINTS WITH SQUARES jxjO(1/e ) [114] [18]
prove negative results for the existence of EPTASs is the fol- DISTINGUISHING SUBSTRING jxjO(1/e ) [115] [93, 116]
lowing connection, observed independently by Bazgan [77] SELECTION
4
and Cesati and Trevisan [78]. CLOSEST SUBSTRING jxjO(1/e ) [85] [88]
PLANAR MPSAT jxjO(1/e ) [91] [92]
PROPOSITION 6.1. If an optimization problem X admits an
PLANAR TMIN jxjO(1/e ) [91] [92]
EPTAS, then the standard parameterization of X is FPT.
PLANAR TMAX jxjO(1/e ) [91] [92]
PLANAR MULTIVALUED MAX 3CSP jxjO(1/e ) [91] [78]
Proof. Assume that we have an algorithm that produces a
(1 þ e )-approximate solution in time f(1/e ) . jxjO(1). Given an
instance (x, k) of the standard parameterization of X, we set two vertices are connected if and only if the two objects
e :¼ 1/(2k), and use the EPTAS to find a (1 þ e )-approximate have nonempty intersection. Intersection graphs of disks, rec-
solution in time f(1/e ) . jxjO(1) ¼ f(2k) . jxjO(1). Assume first tangles, line segments and other objects arise in applications
that X is a minimization problem. If the optimum is at most such as facility location [79], frequency assignment [80] and
k, then the cost of the approximate solution is at most (1 þ map labeling [81]. The first four lines in Table 2 show that
e )k  k þ 1/2 , k þ 1. As the cost is integer, the cost of the if we restrict MAXIMUM INDEPENDENT SET and MINIMUM
approximate solution is at most k. If the optimum is greater DOMINATING SET to the intersection graphs of unit-radius
than k, then the cost of the approximate solution is also disks or unit-size squares, then the problem admits a PTAS
greater than k. Therefore, by checking whether the cost of [76, 82, 83], but does not have an EPTAS [18, 84]. COVERING
the approximate solution is at most k, we can decide POINTS WITH SQUARES is also a geometric problem, but it is
whether the optimum is at most k. The proof is similar in not defined in terms of intersection graphs. In this problem,
the case of maximization problems: if the optimum is at n points are given in the plane, and the task is to cover all of
least k, then the cost of an (1 þ e )-approximate solution is at them with at most k (axis-parallel) unit squares. The squares
least k/(1 þ e ) ¼ k 2 1/(2 þ 1/k) . k 2 1. A can be at arbitrary locations in the plane.
We can use the contrapositive of Proposition 6.1 to show DISTINGUISHING SUBSTRING SELECTION and CLOSEST SUB-
that it is unlikely that a particular problem admits an EPTAS. STRING are two pattern matching problems that are motivated
by applications in computational biology. In the CLOSEST SUB-
COROLLARY 6.1. If the standard parameterization of an STRING problem, the task is to find a string of length L that
optimization problem is W[1]-hard, then the optimization approximately appears in each of the k input strings, where
problem does not have an EPTAS (unless FPT ¼ W[1]). ‘approximately appears’ means that it appears with at most
It has to be remarked that the converse of Proposition 6.1 is d mismatches. The goal is to minimize this number d. In the
not true. For example, MINIMUM VERTEX COVER is FPT, but it DISTINGUISHING SUBSTRING SELECTION problem the input
is APX-hard, hence it does not even have a PTAS. Therefore, strings are divided into ‘good’ strings and ‘bad’ strings, and
Proposition has 6.1 limited applicability. we have to find a string of length L that approximately
appears in each of the good strings, but does not appear
(even approximately) in any of the bad strings. We do not
give a more detailed description of these problems, as many
6.1. Examples
variants, parameterizations and optimization goals are
Table 2 lists problems where the existence of an EPTAS was defined in the literature, see [85 – 89].
ruled out using Proposition 6.1. The first four problems are Khanna et al. [90] defined classes of optimization problems
geometric problems. For a set V of geometric objects, the that have PTASs. The key to approximability in these pro-
intersection graph of V is a graph with vertex set V where blems is the underlying planar structure, which allows us to

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

 PARAMETERIZED COMPLEXITY AND APPROXIMATION ALGORITHMS 71

use Baker’s layering technique [34] and algorithms for Motwani [91] show that MPSAT, TMIN, TMAX, MULTIVA-
bounded tree width. The problems are formulated using LUED MAX kCSP all admit a PTAS if the incidence graph is
Boolean logical expressions. Recall that a formula in disjunc- planar. The running time is jxjO(1/e ), hence it is not an
tive normal form (DNF) is a disjunction of terms, e.g. EPTAS. Cesati and Trevisan [78] have shown that the standard
(x1 ^ x̄3)_(x̄1 ^ x̄2 ^ x̄4) _ x̄3. A DNF is positive (respectively parameterization of PLANAR MULTIVALUED MAX 3CSP is
negative), if every literal is positive (respectively negated). W[1]-hard, hence by Proposition, 6.1 the PTAS cannot be
The weight of an assignment is the number of variables that improved to an EPTAS, unless W[1] ¼ FPT. Similar results

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

are set to true. were obtained for MPSAT, TMIN, TMAX by Cai et al. [92].
THEOREM 6.1 [78, 92]. The planar problems MPSAT, TMIN,
MPSAT TMAX, PLANAR MULTIVALUED MAX 3CSP do not admit an
Input: A collection C ¼ ff1, . . . , fng of DNFs. EPTAS, unless W[1] ¼ FPT.
Find: An assignment c.
Goal: Maximize the number of DNFs satisfied by c.
6.2. Parameterization by 1/e
All the negative results in Table 2 were obtained using Prop-
TMIN osition 6.1. While this method is simple and apparently works,
Input: A collection C ¼ ff1, . . . , fng of positive DNFs. there is something unnatural about it. Our aim is to prove that
Find: An assignment c that satisfies every DNF in C 1/e cannot be taken out of the exponent of the input size. In
Goal: Minimize the weight of c. order to show this, we consider the standard parameterization,
and prove that k cannot be taken out of the exponent. But it
could not be done somehow more directly, by defining 1/e
TMAX to be the parameter? Intuitively, it is clear that what we are
Input: A collection C ¼ ff1, . . . , fng of negative DNFs. trying to determine is whether the problem is FPT, with 1/e
Find: An assignment c that satisfies every DNF in C. being the parameter. However, it is not entirely clear what
Goal: Maximize the weight of c. ‘defining 1/e to be the parameter’ means. The problem is
that the parameterized complexity theory is not developed
for optimization problems, and optimization problems are
These problems generalize many of the standard optimization usually studied via the standard parameterization. In the
problems: for example, it is easy to see that MAX CUT can decision version, however, it is not immediately clear what
be reduced to MPSAT; MAXIMUM INDEPENDENT SET can be the meaning of e is, as we do not have to produce a solution.
reduced to TMAX; and MINIMUM VERTEX COVER can be To make the idea of parameterization by 1/e technically rigor-
reduced to TMIN. (In all three reductions, we associate a vari- ous, we introduce a ‘gap’ version of the standard parameteri-
able with each vertex and a DNF with each edge.) Cesati and zation: the input comes with a promise stating that there is a
Trevisan [78] study a problem of similar flavor, where the gap between the yes-instances and the no-instances. Parame-
variables are not necessarily Boolean, they can have arbitrary terization by 1/e is parameterization by this gap size. If X is
domains. Let x1, . . . , xn be variables and let Di be the domain an optimization problem, then we define
of variable xi. A constraint over x1, . . . , xn is an arbitrary
relation over D1, . . . , Dn, i.e. each possible combination of
values assigned to the variables either satisfies the constraint GAP-X
or not. A constraint is given in the input by listing the combi- Input: An instance x of X, an integer k, and a rational
nation of values that satisfy the constraint. e . 0 such that
8
>
> either OPTðxÞ  k or OPTðxÞ . ð1 þ eÞk
MULTIVALUED MAX kCSP >
>
< if goal ¼ min
Input: A collection C ¼ fC1, . . . , Cmg of constraints; each
k
constraint is on k variables. >
> either OPTðxÞ  k or OPTðxÞ ,
>
> 1þe
Find: An assignment c. :
if goal ¼ max
Goal: Maximize the number of constraints satisfied by c.

Parameter: d1/e e
Given an instance of the above problems, the incidence graph Decide: OPTðxÞ  k; if goal ¼ min;
is a bipartite graph defined by associating a vertex to each vari- OPTðxÞ  k; if goal ¼ max:
able and each clause (constraint), and by connecting each vari-
able to every clause (constraint) where it appears. Khanna and

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

72 D. MARX

An approximation scheme for optimization problem X can

be used to decide GAP-X. Let X be a minimization problem
(the argument for maximization problems is similar). If
(x, k, e ) is a yes-instance of GAP-X, then OPT(x)  k, and
the approximation scheme can produce a (1 þ e )-approximate
solution with cost at most (1 þ e )k. On the other hand, if (x, k,
e ) is a no-instance, then by assumption OPT(x) . (1 þ e )k

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

holds, thus the cost of the solution produced by the approxi-
mation scheme is strictly greater than (1 þ e )k. This means
that deciding an instance (x, k, e ) of GAP-X is not more diffi-
cult than obtaining a (1 þ e )-approximate solution. Therefore,
we have the following proposition.

PROPOSITION 6.2. If optimization problem X admits a PTAS,

then GAP-X is in XP. If X admits an EPTAS, then GAP-X is
in FPT. FIGURE 1. Structure of the instance constructed in Theorem 6.2
(n ¼ 4, t ¼ 3).
Recall that XP is the class of parameterized problems that
can be decided in polynomial time for every fixed value of
the parameter. Proposition 6.2 can be used to show that a
formulas are defined as
problem X does not have an EPTAS: if GAP-X is W[1]-hard,
then X does not have an EPTAS unless FPT ¼ W[1]. This  Wn
ða ^ai;j;s ^bi1;j;s ^bi;j;s Þ; if i ¼ j;
technique is potentially more powerful than using Proposition fi;j ¼ W s¼1 i;j1;s
ða ^a ^b ^b
ðx;yÞ[EðGÞ i;j1;x i;j;x i1;j;y i;j;y Þ; otherwise:
6.1: it might apply even to problems whose standard parame-
terizations are FPT, since k is not a parameter of GAP-X.
To demonstrate the use of Proposition 6.2, we re-prove a This completes the description of the instance. The optimum
result of [92] and show that PLANAR TMIN does not admit is either at most k, or at least k þ 1 . (1 þ e )k ¼ (1 þ 1/
an EPTAS. The reduction is based on the framework of (2k))k ¼ k þ 1/2. Therefore, the gap size requirement is
matrix-type reductions introduced in [84], but the presentation satisfied.
here is self-contained. Assume that G has a clique v1, . . . , vt. For every i,j, we set
ai,j,vi of block Ai,j and bi,j,vj of block Bi,j to true. We claim that
THEOREM 6.2. GAP-PLANAR-TMIN is W[1]-hard. Thus this assignment of weight 2t(t þ 1) satisfies every formula.
PLANAR TMIN does not have an EPTAS, unless FPT ¼ W[1]. Indeed, if i ¼ j ¼ r, then fi,j is satisfied by the term
ai,j21,vr ^ ai,j,vr ^ bi21,j,vr ^ bi,j,vr; if i = j, then fi,j is satisfied
by the term ai,j21,vi ^ ai,j,vi ^ bi21,j,vj ^ bi,j,vj (note that (vi, vj) is
Proof. The reduction is from MAXIMUM CLIQUE. Given a an edge of G, hence fi,j has such a term).
graph G and an integer t, we construct an instance of GAP- For the other direction, suppose that G has no clique of
PLANAR-TMIN with k ¼ 2t(t þ1) and e ¼ 1/(2k) such that size t, but there is a satisfying assignment of weight at most
(1 þ e )k. Notice that (1 þ e )k , k þ 1, hence the weight of
† if G has a clique of size t, then the constructed instance the assignment is at most k. Furthermore, each block of vari-
has a solution with weight at most k; ables has to contain at least one true variable, otherwise the
† if G does not have a clique of size t, then the constructed adjacent formulas could not be satisfied. Therefore, the
instance does not have a solution with weight at most weight of the assignment is exactly k, and there is exactly
(1 þ e )k. one true variable in each block.
For every term of fi,j, there is an s such that the term con-
Assume that G has n vertices f1, 2, . . . , ng. The constructed tains both ai,j21,s and ai,j,s. This means that if fi,j is satisfied,
instance has 2t(t þ 1)n variables. The variables are arranged then the true variables in blocks Ai,j21 and Ai,j have the
into blocks Ai,j (1  i  t, 0  j  t) and Bi,j (0  i  t, 1  same index. Furthermore, this is true for all the blocks in
j  t), where each block contains n variables. The variables row i, hence there is a value vi such that ai,j,vi is true for
in block Ai,j (respectively, Bi,j) will be denoted by ai,j,s every 1  j  t. Similarly, for every column j, there is a
(respectively, bi,j,s) for s ¼ 1, . . . , n. value vj0 such that variable bi,j,vj0 is true. We claim that vi ¼ vi0
We construct t 2 DNFs such that fi,j (1  i, j  t) for every 1  i  t. Indeed, if a term of fi,i is satisfied, then
contains variables only from blocks Ai,j21, Ai,j, Bi21, j, Bi,j. both ai,i,s and bi,i,s have to be true for some s, which implies
As shown in Fig. 1, the incidence graph is planar. The s ¼ vi ¼ vi0. Now the vertices v1, . . . , vt form a clique: if

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

 PARAMETERIZED COMPLEXITY AND APPROXIMATION ALGORITHMS 73

vi and vj are not connected by an edge, then none of the terms parameter exponentially, thus all we can prove is that there
in fi,j are satisfied. This proves the correctness of the is no f(1/e ) . jxjo(log 1/e ) time PTAS.
reduction. A In a similar way, we can obtain lower bounds using
Proposition 6.2. For example, the reduction in Theorem 6.2
constructs instances where the parameter 1/e is O(t 2). There-
fore, we have the following corollary.
6.3. Proving lower bounds pffiffiffiffiffiffiffiffi
1=e )
COROLLARY 6.2. There is no f(1/e ) . jxjoð time PTAS for

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

The fact that an optimization problem does not have an
PLANAR TMIN, unless ETH fails.
EPTAS (unless FPT ¼ W[1]) does not rule out the possibility
of having a PTAS of running time jxjO(log log log 1/e ), for 1/4
This result is somewhat better than the f(1/e ) . jxjo((1/e ) )
example. Such a running time would be almost as good as a
bound that follows from the proof in [92]. However, there is
polynomial-time algorithm. However, by a more careful
still a gap between this lower bound and the jxjO(1/e ) upper
analysis and by having a somewhat stronger complexity-
bound given by Khanna and Motwani [91]. It would be inter-
theoretic assumption, we can actually prove lower bounds
esting to close this gap, possibly by proving stronger lower
on the exponent of jxj.
bounds. Using bidimensionality theory [97, 98], one may
MAXIMUM CLIQUE can be solved in jxjO(k) time by brute
show that the standard parameterizationpffiffiffi of PLANAR TMIN
force. It is believed that this is essentially best possible; in Oð kÞ
can be solved in time f(k) . jxj . This means that no
fact, Chen et al. [93] show that an f(k) . jxjo(k) time algorithm pffiffiffiffiffiffiffiffi
bound better than f(1/e )jxjOð 1=e Þ can be obtained for
would imply that 3SAT can be solved in 2o(n) time. The
PLANAR TMIN using Proposition 6.1. Note that if we are
fastest known algorithm for n-variable 3-SAT runs in random-
using Proposition 6.2, then there is no such apparent barrier.
ized time 1.32216n [94] (improving on the 2n brute force
It seems that we have to abandon Proposition 6.1 and directly
algorithm). It is believed the running time of every algorithm
parameterize by 1/e , if we want to prove tight lower bounds on
must be exponential in n, i.e. no algorithm with running
the efficiency of approximation schemes.
time 2o(n) exists. This assumption is known as the
Exponential Time Hypothesis (ETH) [95], and it is equivalent
to the parameterized complexity conjecture FPT = M[1]
7. CONCLUSIONS
(see [2, 96]).
This lower bound on MAXIMUM CLIQUE can be transfered We have investigated several ways in which parameterized
to other parameterized problems via reductions. Assume that complexity and the theory of approximation algorithms can
a problem Q is proved W[1]-hard by a parameterized benefit from each other. It is not clear at the moment which
reduction from MAXIMUM CLIQUE, and the reduction con- directions will lead to fruitful areas of research. It seems
structs instances such that the parameter is at most p(k) that there are many examples of FPT-approximability for
(for some function p), where k is the parameter of the parameterized instances (as discussed in Section 3), and it
maximum clique instance. (The definition of parameterized can be reasonably expected that many such results will
reduction requires that such a function p(k) exists.) If Q follow. It is much less clear to what extent standard
can be solved in time f(k) . jxjg(k), then the reduction FPT-approximability (Section 4, 5) will be important. The
implies that MAXIMUM CLIQUE can be solved in time f 0 (k) relative lack of examples might be the sign that we are
. jxjg(p(k)). Therefore, the problem Q cannot be solved in missing some fundamental method. Evaluating the efficiency
21
time f00 (k) . jxjo(p (k)) for any function f00 (k), unless ETH of approximation schemes (Section 6) has become a standard
fails. technique, with results of this type steadily accumulating.
Proposition 6.1 shows that if an optimization problem Besides obtaining such results for a wider range of problems,
admits an approximation scheme with running time f(1/e ) . here the challenge is to prove tight results.
jxjg(1/e ), then the standard parameterization can be solved in There are some further connections between parameterized
f(2k) . jxjg(2k) time. Therefore, a lower bound on the standard complexity and approximation algorithms that we have not
parameterization can be used to bound the efficiency of any touched upon. Kernelization of a parameterized problem is a
PTAS for the problem. The W[1]-hardness results in Table 2 polynomial-time preprocessing algorithm that constructs an
are obtained by a parameterized reduction from MAXIMUM equivalent instance such that the size of the new instance
CLIQUE, and these reductions increase the parameter at most can be bounded by a function of the parameter. For
by some polynomial function p(k). Therefore, assuming example, Nemhauser and Trotter [99] present a preprocessing
ETH, for each of these problems there is an integer c such algorithm for the MINIMUM VERTEX COVER problem that
1/c
that there is no f(k) . jxjo(k ) time algorithm for the standard constructs an instance with at most 2k vertices (where k is
parameterization, which implies that there is no f(1/e ) . the parameter of the new instance). In many cases, such a
1/c
jxjo((1/e ) ) time PTAS. The only exception is CLOSEST SUB- linear-size kernelization algorithm can be used to obtain a
STRING: the reduction presented in [88] increases the constant-factor approximation algorithm. Therefore, any

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

74 D. MARX

lower bound on the approximation ratio (obtained by the PCP † Is there a standard FPT-approximation algorithm for
theorems) immediately translates into a lower bound on the MAXIMUM CLIQUE or MINIMUM DOMINATING SET with
kernel size. (A different approach for obtaining kernelization some performance ratio function?
lower bounds is proposed in [100].) † Prove tight lower bounds (assuming ETH) on the effi-
Another interesting connection appears in the use of the ciency of the PTAS for some natural problem
iterative compression technique. The main idea of iterative (MPSAT, TMIN, geometric problems, etc.)
compression is that instead of solving the search problem

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

‘find a solution of size k,’ we solve the possibly easier com-
pression problem ‘given a solution of size k þ 1, find a sol- ACKNOWLEDGEMENTS
ution of size k.’ It turns out that for certain minimization I would like to thank Mike Fellows for asking me to write this
problems if we can solve the compression problem, then we survey, and for the many ideas, comments and references that
can solve the search problem as well by iteratively solving were essential in compiling the material. I am grateful to
larger and larger instances [31, 60, 101– 104]. Furthermore, Martin Grohe and Magdalena Grüber for their comments on
as observed in [104], if we can solve the compression the manuscript.
problem and have a constant-factor approximation algorithm,
then the approximate solution can be compressed, and there is
no need for the iterative trick. In fact, an approximation algor-
REFERENCES
ithm with arbitrary performance ratio function (Section 5)
would be sufficient for this application. [1] Demaine, E.D., Hajiaghayi, M. and Kawarabayashi, K. (2005)
In counting problems, the task is not only to check if there is Algorithmic graph minor theory: decomposition, approxi-
a solution, but we have to determine the total number of sol- mation, and coloring. Proc. 46th Annual IEEE Symp.
utions to the problem. The counting problem can be hard Foundations of Computer Science (FOCS 2005), Pittsburgh,
even for those problems where the existence of a solution is PA, October 23–25, pp. 637–646.
polynomial-time decidable. For example, checking whether [2] Flum, J. and Grohe, M. (2006) Parameterized Complexity
a graph has a perfect matching can be done in polynomial Theory. Springer-Verlag, Berlin.
time, but Valiant [105] has shown that counting the number [3] Downey, R.G. and Fellows, M.R. (1999) Parameterized
of perfect matching is complete for the complexity class #P. complexity. In Monographs in Computer Science.
A similar increase in complexity appears in the case of para- Springer-Verlag, New York.
meterized problems: finding a path of length k is FPT [16, [4] Ausiello, G., Crescenzi, P., Gambosi, G., Kann, V.,
106, 107], but counting the number of these paths is Marchetti-Spaccamela, A. and Protasi, M. (1999) Complexity
#W[1]-hard [108]. Given the hardness of exact counting, it is and Approximation Springer-Verlag, Berlin.
very natural to investigate the possibility of approximate [5] Håstad, J. (2001) Some optimal inapproximability results.
counting. In the parameterized complexity framework, J. ACM, 48, 798–859.
Arvind and Raman [109] present an FPT-time randomized [6] Böckenhauer, H.-J., Hromkovič, J., Kneis, J. and Kupke, J.
(2006) On the parameterized approximability of TSP with
approximation scheme for counting the number of subgraphs
deadlines. Theory Comput. Syst., To appear.
isomorphic to a bounded tree width graph H (thus, in particu-
[7] Har-Peled, S. and Mazumdar, S. (2004) On coresets for
lar, the algorithm can be used to approximate the number of
k-means and k-median clustering. Proc. 36th Annual ACM
paths of length k). Recently, Müller [110] proved a parameter-
Symp. Theory of Computing, Chicaco, IL, pp. 291–300.
ized complexity analog of a classical result of Stockmeyer by
ACM, New York.
showing that approximately counting the number of solutions
[8] Christofides, N. (1976) Worst-case analysis of a new heuristic
for a problem in W[P] can be done in randomized FPT-time for the travelling salesman problem. Technical report 388,
with a W[P] oracle. Graduate School of Industrial Administration,
Finally, to motivate further research, we propose four open Carnegie-Mellon University, Pittsburgh.
questions, related to the four main issues discussed in the [9] Papadimitriou, C. H. and Yannakakis, M. (1993) The traveling
paper. salesman problem with distances one and two. Math. Oper.
Res., 18, 1 –11.
[10] Chen, J., Kanj, I.A. and Jia, W. (2001) Vertex cover: further
† Investigate the approximability of geometric problems, observations and further improvements. J. Algorithms, 41,
parameterized by the dimension. 280–301.
† Find a convincing example where no FPT-time exact [11] Sunil Chandran, L. and Grandoni, F. (2005) Refined
algorithm is known for a problem, but it has a standard memorization for vertex cover. Inf. Process. Lett., 93, 125–131.
FPT-approximation algorithm with constant perform- [12] Khot, S. and Regev, O. (2003) Vertex cover might be hard to
ance ratio. approximate to within 2 2 1. Proc. 18th Annual IEEE

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

 PARAMETERIZED COMPLEXITY AND APPROXIMATION ALGORITHMS 75

Conference on Computational Complexity (CCC’03), [31] Marx, D. and Schlotter, I. (2007) Obtaining a planar graph by
pp. 397 –386. vertex deletion. To appear in WG’07.
[13] Guo, J., Niedermeier, R. and Wernicke, S. (2005) [32] Lichtenstein, D. (1982) Planar formulae and their uses. SIAM
Parameterized complexity of generalized vertex cover J. Comput., 11, 329–343.
problems. Proc. 9th Workshop on Algorithms and Data [33] Lipton, R. J. and Tarjan, R.E. (1980) Applications of a planar
Structures (WADS’05), Lecture Note Series in Computer separator theorem. SIAM J. Comput., 9, 615–627.
Science, Vol. 3608, pp. 36 –48. Springer-Verlag, Berlin. [34] Baker, B.S. (1994) Approximation algorithms for NP-

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

[14] Hochbaum, D.S. and Pathria, A. (1998) Analysis of the greedy complete problems on planar graphs. J. Assoc. Comput.
approach in covering problems. Nav. Res. Q., 45, 615– 627. Mach., 41, 153–180.
[15] Petrank, E. (1994) The hardness of approximation: gap [35] Grohe, M. (2003) Local tree-width, excluded minors, and
location. Comput. Complexity, 4, 133 –157. approximation algorithms. Combinatorica, 23, 613–632.
[16] Alon, N., Yuster, R. and Zwick, U. (1997) Finding and [36] Demaine, E.D. and Hajiaghayi, M. (2004) Equivalence of local
counting given length cycles. Algorithmica, 17, 209– 223. treewidth and linear local treewidth and its algorithmic
[17] Gonzalez, T.F. (1985) Clustering to minimize the maximum applications. Proc. 15th Annual ACM-SIAM Symp. Discrete
intercluster distance. Theoret. Comput. Sci., 38, 293– 306. Algorithms (SODA 2004), New Orleans, LA, January 11 –13,
[18] Marx, D. (2005) Efficient approximation schemes for pp. 833–842.
geometric problems? Proc. 13th Annual European Symp. [37] Bar-Noy, A., Halldórsson, M.M., Kortsarz, G., Salman, R.
Algorithms (ESA 2005), Lecture Notes in Computer Science, and Shachnai, H. (2000) Sum multicoloring of graphs.
Vol. 3669, pp. 448– 459. Springer, Berlin. J. Algorithms, 37, 422–450.
[19] Agarwal, P.K. and Procopiuc, C.M. (2002) Exact and [38] Nicoloso, S., Sarrafzadeh, M. and Song, X. (1999) On the
approximation algorithms for clustering. Algorithmica, 33, sum coloring problem on interval graphs. Algorithmica, 23,
201 –226. 109–126.
[20] Trevisan, L. (2000) When Hamming meets Euclid: the [39] Jansen, K. (1997) The optimum cost chromatic partition
approximability of geometric TSP and Steiner tree. SIAM problem. Proc. Algorithms and complexity, Rome, Lecture
J. Comput., 30, 475 –485. Notes in Computer Science, Vol. 1203, pp. 25–36. Springer,
[21] Cabello, S., Giannopoulos, P. and Knauer, C. (2006) On the Berlin.
parameterized complexity of d-dimensional point set pattern [40] Giaro, K. and Kubale, M. (2000) Edge-chromatic sum of
matching. Proc. Int. Workshop on Parameterized and Exact trees and bounded cyclicity graphs. Inf. Process. Lett., 75,
Computation (IWPEC 2006), Lecture Notes in Computer 65 –69.
Science, Vol. 4169, pp. 175 –183. [41] Salavatipour, M.R. (2003) On sum coloring of graphs. Discret.
[22] Bodlaender, H.L. (1993) A tourist guide through treewidth. Appl. Math., 127, 477–488.
Acta Cybern., 11, 1 –21. [42] Marx, D. (2004) Complexity results for minimum sum edge
[23] Diestel, R. (2005) Graph Theory. Graduate Texts in coloring. Manuscript.
Mathematics, (3rd edn), Springer-Verlag, Berlin. [43] Marx, D. (2005) Minimum sum multicoloring on the edges of
[24] Cai, L. (2003) Parameterized complexity of vertex colouring. planar graphs and partial k-trees. Proc. 2nd Int. Workshop on
Discret. Appl. Math., 127, 415 –429. Approximation and Online Algorithms (WAOA 2004),
[25] Guo, J., Hüffner F. and Niedermeier, R. (2004) A structural Lecture Notes in Computer Science, Vol. 3351, pp. 9–22.
view on parameterizing problems: distance from triviality. Springer, Berlin.
Proc. Int. Workshop on Parameterized and Exact [44] Cai, L. and Huang, X. (2006) Fixed-parameter approximation:
Computation (IWPEC 2006), Lecture Notes in Computer conceptual framework and approximability results. Proc. Int.
Science, Vol. 3162, pp. 162 –173. Springer, Berlin. Workshop on Parameterized and Exact Computation
[26] Marx, D. (2006) Parameterized coloring problems on chordal (IWPEC 2006), Vol. 4169, Lecture Notes in Computer
graphs. Theoret. Comput. Sci., 351, 407 –424. Science, pp. 96 –108.
[27] Feige, U. and Kilian, J. (1998) Zero knowledge and the [45] Chen, Y., Grohe, M. and Grüber, M. (2006) On parameterized
chromatic number. J. Comput. Syst. Sci., 57, 187– 199. approximability. Proc. Int. Workshop on Parameterized and
[28] Stockmeyer, L. (1973) Planar 3-colorability is NP-complete. Exact Computation (IWPEC 2006), Lecture Notes in
SIGACT News, 5: 19 –25. Computer Science, Vol. 4169, pp. 109– 120.
[29] Appel, K. and Haken, W. (1989) Every Planar Map is Four [46] Downey, R., Fellows, M. and McCartin, C. (2006)
Colorable. Contemporary Mathematics, Vol. 98 (American Parameterized approximation algorithms. Proc. Int.
Mathematical Society, Providence, RI), (With the Workshop on Parameterized and Exact Computation
collaboration of J. Koch). (IWPEC 2006), Lecture Notes in Computer Science, Vol.
[30] Robertson, N., Sanders, D.P., Seymour, P. and Thomas, R. (1996) 4169, pp. 121–129.
Efficiently four-coloring planar graphs. Proc. Twenty-eighth [47] Bodlaender, H.L. (1996) A linear-time algorithm for finding
Annual ACM Symposium on the Theory of Computing, tree-decompositions of small treewidth. SIAM J. Comput.,
Philadelphia, PA, 1996, pp. 571–575. ACM, New York. 25, 1305– 1317.

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

76 D. MARX

[48] Robertson, N. and Seymour, P.D. (1995) Graph minors. XIII. [65] Dı́az, J., Petit, J. and Serna, M. (2002) A survey of graph layout
The disjoint paths problem. J. Comb. Theory Ser. B, 63, problems. ACM Comput. Surv., 34, 313–356.
65 –110. [66] Bodlaender, H.L., Fellows, M.R. and Hallett, M.T. (1994)
[49] Reed, B.A. (1992) Finding approximate separators and Beyond NP-completeness for problems of bounded width
computing tree width quickly. STOC’92: Proc. (extended abstract): hardness for the W hierarchy. In STOC
Twenty-fourth Annual ACM Symp. Theory of Computing, ’94: Proc. Twenty-sixth Annual ACM Symp. Theory of
New York, NY, USA, pp. 221 –228. ACM Press. computing, New York, NY, USA, pp. 449–458. ACM Press.

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

[50] Reed, B.A. (1997) Tree width and tangles: a new connectivity [67] Thilikos, D.M., Serna, M.J. and Bodlaender, H.L. (2000)
measure and some applications. Surveys in combinatorics, 1997 Constructive linear time algorithms for small cutwidth and
(London), London Mathematical Society Lecture Note Series, carving-width. Proc. Algorithms and Computation, Taipei,
Vol. 241, pp. 87–162. Cambridge Univ. Press, Cambridge. Lecture Notes in Computer Science, Vol. 1969, pp. 192–
[51] Amir, E. (2001) Efficient approximation for triangulation of 203. Springer, Berlin.
minimum treewidth. In UAI ’01: Proc. 17th Conf. [68] Bodlaender, H.L. Unpublished result.
Uncertainty in Artificial Intelligence, San Francisco, CA, [69] Fellows, M.R. Unpublished result.
USA, pp. 7– 15. Morgan Kaufmann Publishers Inc. [70] Courcelle, B., Engelfriet, J. and Rozenberg, G. (1991)
[52] Courcelle, B. and Oum, S.-I. (2007) Vertex-minors, monadic Contextfree handle-rewriting hypergraph grammars. Graph
second-order logic, and a conjecture by Seese. J. Comb. Grammars and Their Application to Computer Science
Theory Ser. B, 97, 91– 126. (Bremen, 1990), Lecture Notes in Computer Science, Vol.
[53] Hliněný, P. (2005) A parametrized algorithm for matroid 532, pp. 253–268. Springer, Berlin.
branch-width. SIAM J. Comput., 35, 259 –277. [71] Fellows, M.R., Rosamond, F.A., Rotics, U. and Szeider, S.
[54] Oum, S.-I. and Seymour, P. (2006) Approximating (2006) Clique-width minimization is NP-hard. Proc. 38th
cliquewidth and branch-width. J. Comb. Theory Ser. B, 96, Annual ACM Sym. Theory of Computing, Seattle, WA, USA,
514 –528. pp. 354–362. ACM press.
[55] Oum, S.-I. (2005) Approximating rank-width and cliquewidth [72] Downey, R.G. (2003) Parameterized complexity for the
quickly. Proc. 31st Int. Workshop on Graph-Theoretic skeptic. Proc. 18th IEEE Annual Conf. Computational
Concepts in Computer Science, pp. 49–58. Complexity, pp. 147– 169.
[56] Grohe, M. and Grüber, M. (2007) Parameterized approxi- [73] Arora, S. (1996) Polynomial time approximation schemes for
mability of the disjoint cycle problem, ICALP ’07, To appear. Euclidean TSP and other geometric problems. Proc. 39th
[57] Bodlaender, H.L., Gilbert, J.R., Hafsteinsson, H. and Kloks, T. Annual Symp. Foundations of Computer Science, pp. 2– 11.
(1995) Approximating treewidth, pathwidth, frontsize, and IEEE Comput. Soc. Press.
shortest elimination tree. J. Algorithms, 18, 238–255. [74] Arora, S. (1998) Polynomial time approximation schemes for
[58] Feige, U., Hajiaghayi, M.T. and Lee, J.R. (2005) Improved Euclidean traveling salesman and other geometric problems.
approximation algorithms for minimum-weight vertex J. ACM, 45, 753–782.
separators. STOC’05: Proc. 37th Annual ACM Symp. Theory [75] Arora, S., Raghavan, P. and Rao, S. (1999) Approximation
of Computing, pp. 563– 572. ACM, New York. schemes for Euclidean k-medians and related problems. In
[59] Bouchitté, V., Kratsch, D., Müller, H. and Todinca, I. (2004) STOC ’98, Dallas, TX, pp. 106–113. ACM, New York.
On treewidth approximations. Discret. Appl. Math., 136, [76] Hunt, III, H.B., Marathe, M.V., Radhakrishnan, V., Ravi, S.S.,
183 –196; The 1st Cologne-Twente Workshop on Graphs Rosenkrantz, D.J. and Stearns, R.E. (1998) NC-approximation
and Combinatorial Optimization (CTW 2001). schemes for NP- and PSPACE-hard problems for geometric
[60] Dehne, F., Fellows, M., Langston, M.A., Rosamond, F. graphs. J. Algorithms, 26, 238–274.
and Stevens, K. (2005) An O(2O(k)n 3) FPT algorithm for the [77] Bazgan, C. (1995) Schémas d’approximation et complexité
undirected feedback vertex set problem. Proc. Computing paramétrée. Technical report, Université Paris Sud.
and Combinatorics, Lecture Notes in Computer Science, [78] Cesati, M. and Trevisan, L. (1997) On the efficiency of
Vol. 3595, pp. 859 –869. Springer, Berlin. polynomial time approximation schemes. Inf. Process. Lett.,
[61] Even, G., Naor, J., Schieber, B. and Sudan, M. (1998) 64, 165–171.
Approximating minimum feedback sets and multicuts in [79] Wang, D.W. and Kuo, Y.-S. (1988) A study on two geometric
directed graphs. Algorithmica, 20, 151– 174. location problems. Inf. Process. Lett., 28, 281–286.
[62] Seymour, P.D. (1995) Packing directed circuits fractionally. [80] Malesińska, E. (1997) Graph-theoretical models for frequency
Combinatorica, 15, 281 –288. assignment problems. PhD Thesis, Technical University of
[63] Slivkins, A. (2003) Parameterized tractability of edge-disjoint Berlin.
paths on directed acyclic graphs. Proc. Algorithms—ESA [81] Agarwal, P.K., van Kreveld, M. and Suri, S. (1998) Label
2003, 11th Annual Eur. Symp. Lecture Notes in Computer placement by maximum independent set in rectangles.
Science, Vol. 2832, pp. 482 –493. Springer, Berlin. Comput. Geom., 11, 209– 218.
[64] Reed, B., Robertson, N., Seymour P. and Thomas, R. (1996) [82] Nieberg, T. and Hurink, J. (2006) A PTAS for the minimum
Packing directed circuits. Combinatorica, 16, 535–554. dominating set problem in unit disk graphs. Proc. 3rd Int.

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

 PARAMETERIZED COMPLEXITY AND APPROXIMATION ALGORITHMS 77

Workshop on Approximation and Online Algorithms (WAOA Lecture Notes in Computer Science, Vol. 3383, Harlem,
2005), Lecture Notes in Computer Science, Vol. 3879, New York, September 29 –October 2, pp. 517–533.
pp. 296 –306. Springer, Berlin. [98] Demaine, E.D., Fomin, F.V., Hajiaghayi, M. and Thilikos,
[83] Nieberg, T., Hurink, J. and Kern, W. (2003) New PTAS for the D.M. (2005) Subexponential parameterized algorithms on
independent set problem in unit disk graphs. Memorandum no. graphs of bounded-genus and H-minor-free graphs. J. ACM,
1688, Department of Applied Mathematics, Universiteit 52, 866–893.
Twente. [99] Nemhauser, G.L. and Trotter, Jr, L.E. (1975) Vertex packings:

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

[84] Marx, D. (2006) Parameterized complexity of independence structural properties and algorithms. Math. Program., 8, 232–
and domination on geometric graphs. Proc. Int. Workshop on 248.
Parameterized and Exact Computation (IWPEC 2006), Lecture [100] Chen, J., Fernau, H., Kanj, I.A. and Xia, G. (2005) Parametric
Notes in Computer Science, Vol. 4169, pp. 154–165. duality and kernelization: lower bounds and upper bounds on
[85] Li, M., Ma, B. and Wang, L. (2002) On the closest string and kernel size. STACS 2005, Lecture Notes in Computer
substring problems. J. ACM, 49, 157 –171. Science, Vol. 3404, pp. 269–280. Springer, Berlin.
[86] Gramm, J., Niedermeier, R. and Rossmanith, P. (2003) [101] Reed, B., Smith, K. and Vetta, A. (2004) Finding odd cycle
Fixedparameter algorithms for closest string and related transversals. Oper. Res. Lett., 32, 299–301.
problems. Algorithmica, 37, 25 –42. [102] Marx, D. (2006) Chordal deletion is fixed-parameter tractable.
[87] Gramm, J., Niedermeier, R. and Rossmanith, P. (2001) Exact In Graph-theoretic concepts in computer science (WG 2006),
solutions for closest string and related problems. ISAAC ’01: Lecture Notes in Computer Science, Vol. 4271, pp. 37– 48.
Proc. 12th Int. Symp. Algorithms and Computation, Lecture [103] Guo, J., Gramm, J., Hüffner, F., Niedermeier, R. and Wernicke,
Notes in Computer Science, Vol. 2223, pp. 441– 453. S. (2005) Improved fixed-parameter algorithms for two feedback
Springer, Berlin. set problems. Proc. 9th Workshop on Algorithms and Data
[88] Marx, D. (2005) The closest substring problem with small Structures (WADS’05), Lecture Note Series in Computer
distances. Proc. 46th Annual IEEE Symp. Foundations of Science, Vol. 3608, pp. 158–168. Springer-Verlag, Berlin.
Computer Science (FOCS’05), Pittsburgh, PA, pp. 63–72. [104] Guo, J., Gramm, J., Hüffner, F., Niedermeier, R.
[89] Fellows, M.R. Gramm J. and Niedermeier, R. (2006) On the and Wernicke, S. (2006) Compression-based fixed-parameter
parameterized intractability of motif search problems. algorithms for feedback vertex set and edge bipartization.
Combinatorica, 26, 141– 167. J. Comput. Syst. Sci., 72, 1386 –1396.
[90] Khanna, S., Sudan, M., Trevisan, L. and Williamson, D.P. [105] Valiant, L.G. (1979) The complexity of computing the
(2001) The approximability of constraint satisfaction permanent. Theoret. Comput. Sci., 8, 189–201.
problems. SIAM J. Comput., 30, 1863–1920. [106] Monien, B. (1985) How to find long paths efficiently. Proc
[91] Khanna, S. and Motwani, R. (1996) Towards a syntactic Analysis and design of algorithms for combinatorial
characterization of PTAS. Proc. Twenty-eighth Annual ACM problems (Udine, 1982), North-Holland Mathamatical
Symp. the Theory of Computing, Philadelphia, PA, 1996, pp. Studies, Vol. 109, pp. 239–254. North-Holland, Amsterdam.
329 –337. ACM, New York. [107] Plehn, J. and Voigt, B. (1991) Finding minimally weighted
[92] Cai, L., Fellows, M., Juedes, D. and Rosamond, F. (2006) The subgraphs. Proc. Graph-theoretic concepts in computer
complexity of polynomial-time approximation. Theory science (Berlin, 1990), Lecture Notes in Computer Science,
Comput. Syst., To appear. Vol. 484, pp. 18–29. Springer, Berlin.
[93] Chen, J., Huang, X., Kanj, I.A. and Xia, G. (2004) Linear FPT [108] Flum, J. and Grohe, M. (2004) The parameterized complexity
reductions and computational lower bounds. Proc. 36th of counting problems. SIAM J. Comput., 33, 892– 922.
Annual ACM Symp. Theory of Computing, Chicago, IL, [109] Arvind, V. and Raman, V. (2002) Approximation algorithms for
pp. 212 –221. ACM, New York. some parameterized counting problems. In ISAAC ’02: Proc.
[94] Rolf, D. (2005) Improved bound for the PPSZ/Schöning- 13th Int. Symp. Algorithms and Computation, Lecture Notes in
algorithm for 3-SAT. Technical report TR05-159, Electronic Computer Science, Vol. 2518, pp. 453–464. Springer, Berlin.
Colloq. on Computational Complexity. [110] Müller, M. (2006) Randomized approximations of
[95] Impagliazzo, R., Paturi, R. and Zane, F. (2001) Which parameterized counting problems. Proc. Int. Workshop on
problems have strongly exponential complexity? J. Comput. Parameterized and Exact Computation (IWPEC 2006),
Syst. Sci., 63, 512 –530. Special issue on FOCS 98 (Palo Lecture Notes in Computer Science, Vol. 4169, pp. 50– 59.
Alto, CA). [111] Chekuri, C. and Khanna, S. (2000) A PTAS for the multiple
[96] Downey, R., Estivill-Castro, V., Fellows, M., Prieto, E. knapsack problem. Proc. Eleventh Annual ACM-SIAM Symp.
and Rosamund, F. (2003) Cutting up is hard to do. In Discrete Algorithms, San Francisco, CA, 2000, pp. 213–222.
Harland, J. (ed.), Electronic Notes in Theoretical Computer ACM, New York.
Science, Vol. 78. Elsevier. [112] Shamir, R. and Tsur, D. (1998) The maximum subforest
[97] Demaine, E.D. and Hajiaghayi, M. (2004) Fast algorithms for problem: approximation and exact algorithms (extended
hard graph problems: bidimensionality, minors, and local abstract). Proc. Ninth Annual ACM-SIAM Symp. Discrete
treewidth. Proc. 12th Int. Symp. Graph Drawing (GD 2004), Algorithms, San Francisco, CA, pp. 394–399. ACM, New York.

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008

78 D. MARX

[113] Erlebach, T., Jansen, K. and Seidel, E. (2005) Polynomial-time Int. Colloquium on Automata, Languages and Programming,
approximation schemes for geometric intersection graphs. Lecture Notes in Computer Science, Vol. 2380, pp. 740–
SIAM J. Comput., 34, 1302– 1323. 751. Springer, Berlin.
[114] Hochbaum, D.S. and Maass, W. (1985) Approximation [116] Gramm, J., Guo, J. and Niedermeier, R. (2003) On exact and
schemes for covering and packing problems in image approximation algorithms for distinguishing substring
processing and VLSI. J. ACM, 32, 130–136. selection. Proc. Fundamentals of Computation Theory,
[115] Deng, X., Li, G., Li, Z., Ma, B. and Wang, L. (2002) A PTAS Lecture Notes in Computer Science, Vol. 2751, pp. 195–
for distinguishing (sub)string selection. ICALP 02: Proc. 29th 209. Springer, Berlin.

Downloaded from https://academic.oup.com/comjnl/article-abstract/51/1/60/504387 by Shandong University user on 19 December 2019

THE COMPUTER JOURNAL, Vol. 51 No. 1, 2008