Approximation Schemesfor Covering and Packing

Problems in Image Processingand VLSI

DORIT S. HOCHBAUM AND WOLFGANG MAASS

University of California. Berkeley, California

Abstract. A unified and powerful approach is presented for devising polynomial approximation schemes
for many strongly NP-complete problems. Such schemesconsist of families of approximation algorithms
for each desired performance bound on the relative error c > 0, with running time that is polynomial
when c is fixed. Thougb the polynomiality of these algorithms depends on the degree of approximation
e being fixed, they cannot be improved, owing to a negative result stating that there are no fully
polynomial approximation schemes for strongly NP-complete problems unless NP = P.
The unified technique that is introduced here, referred to as the shifting strategy, is applicable to
numerous geometric covering and packing problems. The method of using the technique and how it
varies with problem parameters are illustrated. A similar technique, independently devised by B. S.
Baker, was shown to be applicable for covering and packing problems on planar graphs.

Categories and Subject Descriptors: F.2.2 [Analysis of Algorithms and Problem Complexity]: Nonnu-
merical Algorithms and Problems-geometricalproblems and computations
General Terms: Algorithms, Theory
Additional Key Words and Phrases: Covering, packing, covering points in the Euclidean space, image
processing, VLSI, shifting strategy, worst case analysis of heuristics, polynomial approximation scheme

1. Introduction
Polynomial approximation schemesare described in this paper for severalstrongly
NP-complete problems that have important applications in the areas of robot
motion planning, VLSI design, image processing, and location. These problems
appear in the contexts of covering and packing with convex objects. One of them
is the square packing problem, which comes up in the attempt to increaseyield in
VLSI chip manufacture. For example, 64K RAM chips, some of which may be
defective, are available on a rectilinear grid placed on a silicon wafer. 2 x 2 arrays
of such nondefective chips could be wired together to produce 256K RAM chips.
In order to maximize yield, we want to pack a maximal number of such 2 X 2
arrays into the array of working chips on a wafer. (Seethe result of Berman et al.
[2], reviewed by Johnson [6], and the NP-completenessresult of Fowler et al. [3].)

The research of the first author was supported in part by the National Science Foundation under grant
ECS 82-04695. During the preparation of this paper, the second author was supported by the Heisenberg
Programm der Deutschen Forschungsgemeinschaft, West Germany.
Authors’ address: D. S. Hochbaum, School of Business Administration, 350 Barrows Hail, University
of California at Berkeley, Berkeley, CA 94720, W. Maass, Department of Mathematics, University of
Illinois, Chicago, IL 60680.
Permission to copy wit.hout fee all or part of this material is granted provided that the copies are not
made or distributed for direct commercial advantage, the ACM copyright notice and the title of the
publication and its date appear, and notice is given that copying is by permission of the Association for
Computing Machinery. To copy otherwise, or to republish, requires a fee and/or specific permission.
0 1985 ACM 0004~5411/85/0100-0130 $00.75

Journal ofthe Association for Computing Machinery. Vol. 32, No. I, January 1985, pp. 130-136.
Approximation Schemesfor Covering and Packing Problems 131
Another problem is covering with disks, that is, given points in the plane, to
identify a minimally sized set of disks (of prescribed radius) covering all points.
One of its applications is in the area of locating emergency facilities such that all
potential customers will be within a reasonably small radius around the facility.
(The complexity results for this problem are reviewed in [6].)
A third problem considered is covering with squares or (rectangles),which has
an important application to image processing,discussedin Tanimoto and Fowler
[7]. Here one wants to store information in square “patches” such that all points
with information (“pixels”) are contained in at leastone of the patches.The general
problem can be describedas:Given points in a Euclidean space(in this application,
on a grid), find a minimally sized set of squaresof prescribedsizecovering all those
points.
None of the above problems was reported to have a bounded error ratio
approximation algorithm. We shall call an algorithm a b-approximation, 6 > 0, for
a certain problem if the error of the value of the solution delivered by the algorithm
divided by the value of the optimal solution does not exceed 6. Obviously, we
would like to identify a &approximation algorithm such that 6 is as small as
possible. In some casesone can specify a family of algorithms such that for each
E > 0 there is an c-approximation algorithm in the family that solves a given
problem instance within relative error 6. Such a family is called an approximation
scheme.The running time of an c-approximation algorithm will increase mono-
tonically with l/t. If the functional dependenceof the running time on the size of
the input and I/C is polynomial, then the schemeis said to be fully polynomial; if,
on the other hand, it is polynomial only in the input size, the scheme is called
polynomial.
All the problems described and consequently their extensionsare NP-complete
in the strong sense(the reader is referred to Garey and Johnson’s [4] comprehensive
review of this concept). As such, there are no fully polynomial approximation
schemesfor theseproblems, unlessNP = P ([4, theorem 6.81).This negativeresult,
however, does not exclude the existence of a polynomial approximation scheme
for these problems, that is, a family of algorithms such that for any specified
relative error E > 0 there is an c-approximation algorithm in the scheme that is
polynomial. Though such schemesare conceptually feasible, their existence has
rarely been reported.
Our main results are the construction of polynomial approximation schemesfor
the above problems which, given the negative result above, are the best possible
results of this type. We present a unified methodology that is helpful for numerous
geometric covering and packing problems and could potentially be applicable to
problems beyond this context. We call this fundamental technique the shifting
strategy and outline the necessaryconditions for its applicability. A similar tech-
nique has been independently discoveredby Baker [I]. Her technique is applicable
to planar graphs, whereasours applies to problems defined in Euclidean space;but
our concepts of strips of bounded width and shifting are analogousto the concepts
of bounded outerplanarity and shifting used by Baker.
The shifting strategy is described and the “shifting lemma” is proved in Section
2. We then proceed in Section 3 with its application to problems of covering with
balls and derive a polynomial approximation scheme for this case. We further
explain the behavior of such schemes for problems of covering with arbitrarily
shapedobjects in higher dimensional spaces.In Section 4 we describethe applica-
tion to packing problems with convex objects, and we summarize our results in
Section 5.
132 D. S. HOCHBAUM AND W. MAAS

Throughout the paper the following notation will be used. ZA denotes the value
of the solution delivered by algorithm A. An optimal solution set is denoted by
OPT and its size b:y ] OPT 1.

2. The Shifting Stra:tegy

The shifting strategyallows us to bound the error of the simple divide-and-conquer
approach by applying it repetitively and selectingthe single most favorableresulting
solution. This approach was successfully used for problems of covering with
nonconvex objects as well [5]. In this section, however, we illustrate the application
of the shifting strategy for covering with planar balls (disks).
Let the set iV of the n given points in the plane be enclosed in an area I. The
goal is to cover these points with a minimal number of disks of diameter D. Let
the shifting parameter be 1.In the first phase the area I is subdivided into vertical
strips of width D, where each strip is left closed and right open. Groups of I
consecutive strips, resulting in strips of width 1-D each, are considered. For any
fixed subdivision of I into strips of width D, there are I different ways of partitioning
I into strips of width 1-D. These partitions can be ordered such that each can be
derived from the previous one by shifting it to the right over distance D. Repeating
the shift 1times we end up with the samepartition we started from. We denote the
1distinct shift partitions that result by &, S2, . . . , S,.
Let A be any algorithm that delivers a solution in any strip of width 1-D (or less).
For a given partition Siplet A(Si) be the algorithm that applies algorithm A to each
strip in the partition Si and outputs the union of all disks used. Such a set of disks
is clearly a feasible solution to the global problem defined on I. This process of
finding a global solution is repeated for each partition Si, i = 1,2, . . . , 1.The shift
algorithm SA, defined for a given local algorithm A, delivers the set of disks of
minimum cardinality among the 1setsdelivered by A(S), . . . , A(S).
Let the performance ratio of an algorithm B be denoted by rE, that is, rB is
defined as the supremum of Z’/] OPT ] over all problem instances.
LEMMA 2.1 (THE SHIFTING LEMMA).

rsA5 rA
( ) 1 + -
1
1
whereA is a local algorithm and 1 is the shifting parameter.
, (2.1)

PROOF. We produce an upper bound on the sum of errors caused by all

algorithms A(Si) for i = 1,2, . . . , 1.
By the definition of rA we have
ZA(S,)5 rAeJzs. 1OPTJ ] , (2.2)
I
where J runs over all strips in partition Si and ] OPTJ ] is the number of disks in
an optimal cover of the points in strip J.
Let OPT be the set of disks in an optimal solution and OPT”‘, . . . , OPT”’ the
set of disks in OPT covering points in two adjacent 1.D strips in the 1, 2, . . . , 1
shifts, respectively.It can easily be seenthat
c. IOPT,] 5 IOPT] + IOPT@]. (2.3)
JinSi

There can be no disk in the set OPT that covers points in two adjacent strips in
more than one shift partition. Therefore, the sets OPT”‘, . . . , OPT(‘) are disjoint
Approximation Schemesfor Covering and Packing Problems 133
and can add up to OPT at most. It follows that

ii1 (IOPT + lOPT’“1) I (, + l).lOPTl. (2.4)

Expressions (2.3) and (2.4) imply

min 2 IOPTJI ~~~I(JIsilOP’Gl)~(l +i) IOfll- (2.5)

i=l,...,/ fins,

Combining inequality (2.5) with (2.2), we obtain

1
2’~ = min ZA(‘J 5 rA. 1 + - .lOPTl, (2.6)
i=l,...,l ( 1)
which establishes (2.1). Q.E.D.
The local algorithm A may itself be derived from an application of the shifting
strategy in lower dimensional space. Repetitive applications of this type yield an
approximation scheme as described in the following section.

3. Polynomial Approximation Schemesfor Covering Problems in Arbitrary

Dimensions
We first illustrate in Theorem 3.1 the method of repetitive applications of the
shifting for the problem of covering with balls in a d-dimensional space. The
remainder of this section consists of a generalization of this concept for other
convex objects.
THEOREM 3.1. Let d 2 1 be somefinite dimension. Then there is a polynomial
time approximation schemeHd such that for every given natural number 12 1, the
algorithm Hi’ delivers a cover of n given points in a d-dimensional Euclidean space
by d-dimensional balls of given diameter D in O(ld(l. 4)d.(2n)d(tJa)d+‘) steps with
performance ratio 5 (1 + l/l)4
PROOF. The considered problem is NP-complete only for d > 1. For d = 1 one
can actually compute an optimal solution in linear time with the following
algorithm: We always place the next interval (i.e., l-dimensional ball) with its left
end at the leftmost point that is not yet covered.
For d = 2 and fixed 1L 1 we use two nested applications of the shifting strategy
from Section 2. We first cut the plane into vertical strips of width 1.D. Then, in
order to cover the points in such a strip, we apply the shifting strategy to the other
dimension. Thus, we cut the considered strip into squares of side length I.D. We
find optimal coverings of points in such a square by exhaustive search. With
(1.J2)2 = 212disks of diameter D we can cover an 1.D x 1.D square compactly;
thus we never need to consider more disks for one square. Further, we can assume
that any disk that covers at least two of the given points has two of these points on
its border. (For disks that cover only one point the following estimate holds
trivially.) Since there are only two ways to draw-a circle of given diameter through
two given points, we only have to consider 2. (2”)possible disk positions, where fi
is the number of given points in the considered square. Thus we have to check at
most O(fi2”‘Ja2) arrangements of disks. We specify the position of each disk by its
center. In order to check whether an arrangement of disks is a feasible cover of the
vi points in the square, we need to determine for each point whether it is within
a distance of at most D/2 from one of the centers. Such a check will require
 134 D. S. HOCHBAUM AND W. MAASS

O(l* . fi) steps,with the assumption that we can determine in one step the distance
betweentwo points on the plane or the center of a disk with the necessaryprecision.
(Note that it is possibleto speedup that processby sorting the disk centersin O(l*)
prespecified subsquares of the region and then determining in which of the
subsquaresthere could be the potential covering disk by binary search in time
O(log I).) The two nested applications of the shifting strategy add another factor 1’
to our global time bound.
For d > 2 one proceeds analogously with d nested applications of the shifting
strategy. Q.E.D.
We have considered in Theorem 3.1 the problem of covering given points with
a minimal number of balls of given size. The method of Theorem 3.1 can easily
be generalizedto y:ieldapproximation schemesfor problems where one coverswith
objects other than balls. For a fixed type of object (of arbitrary fixed shape) we
define D as the maximum diameter of such an object. In a manner similar to
Theorem 3.1 we cut the considered d-dimensional spacein a number of different
ways (“shifting”) into d-dimensional cubeswith sidesof length 1-D. One can always
find a local algorithm that proceeds by enumeration in the same way as the
algorithm for balls in Theorem 3.1. But now the number of objectsof the considered
type that are needed to cover a d-dimensional cube with sides of length 1.D will
depend on the ratio between D and the maximal d such that a d-dimensional cube
with sidesof length b is contained in a covering object of the consideredtype. The
running time of the resulting approximation algorithm Hj’ will depend exponen-
tially on this ratio D/b. For instance, for objects with known orientation the
expression(I. a)d in the exponent will be replaced by (I-o/fi)d. This ratio O/d is
usually of interest also in other contexts. Note that, e.g., for rectangles of
size a x b in two dimensions this ratio is closely related to the “aspect ratio”
max(a/b, b/a). We have shown in another paper [5] that in at least one important
case one can eliminate the ratio o/a from the exponent of the running time by
replacing the local enumeration algorithm by another approximation scheme.
In certain applications the covering problem is defined in terms of objects with
fixed orientation. This is the case, for instance, with the covering with squares
problem in the context of image processing[71.This additional constraint simplifies
the problem in that the trick illustrated in the following corollary often sufficesto
eliminate O/b from the exponent of the running time.
COROLLARY 3.2.. Consider the problem of covering n given points in d-space
with a minimal number of rectilinear blocks (the sides of which have given lengths
D,, . . . , Dd) oriented with sides parallel to the axes. There is a polynomial-time
approximation scheme Hd such that for every given integer 1 2 1, the algorithm
H;’ delivers a cover in O(ld. n21d+‘)steps with peformance ratio d (1 + l/l)d.
This corollary is proved in the same way as Theorem 3. i, except that the cuts
orthogonal to the I’th axis are introduced at a distance 1.Di from each other.

4. Application of the Technique to Packing Problems

In a packing problem one wants to place without overlap a maximal number of
objects of given shape with a given area. Since the error analysis of the shifting
strategy remains true for such problems, we can use algorithms similar to those in
Section 3. We consider asan example the problem of packing with squaresdiscussed
in the introduction. The squaresin this case have to be placed so that their sides
Approximation Schemesfor Covering and Packing Problems 135
coincide with the lines of an overlaying rectilinear grid. The following theorem can
also be generalizedto packing problems without such a restriction.
THEOREM 4.1. There is a polynomial-time approximation schemefor the prob-
lem of packing a maximal number of k x k squaresCfora natural number k) into
an area that is given by n squares of unit size on a rectilinear grid. The approxi-
mation algorithm with parameter 1has an error ratio 5 (1 + 1/l)2 and runs in time
O(k2.12-n’2).
PROOF. One reduces the problem via two nested applications of the shifting
strategy to a local packing problem in an 1-k x 1.k square. The local packing
problem can be solved by enumeration in O(fi’) steps,where fi is the number of
given unit squaresthat fall into the considered 1.k x I. k square. We assumehere
that a list of all possibleplacements of k x k squaresthat lie totally inside the area
of the n given unit squareshas been generatedas part of the input.
The shifting strategy that has been described in Section 2 for the covering
problem can be translated to packing problems with the help of the following
observation. The absolute error that results from solving the packing problem
separatelyin each strip for a given shift can be bounded by the number of k x k
squares in an optimal global solution of the packing problems that include unit
squaresin two different strips. Q.E.D.
For fixed 1one has the option of speedingup the corresponding approximation
algorithm by using some faster algorithm for optimal packings in 1.k x 1.k squares
(see,e.g., the analogous approach in [5, sec. 51).
Remark. We also get polynomial approximation schemesfor many packing
problems in higher dimensions, with arbitrary orientations and with other objects
than squares-see the analogous discussion in Section 3.

5. Conclusion
All the approximation schemesdescribed in this paper use a local algorithm that
solves the covering or packing problem in a cube with sides of length 1.D. An
interesting direction for future researchis to find optimal local algorithms that are
more efficient than enumeration. Another alternative is to consider for such local
problems a heuristic that belongs to some approximation scheme.Together with
the shifting strategydefined in this paper such a heuristic, if extant, will produce a
faster overall approximation algorithm.
The results of this paper and Baker’s paper [I] raise the question of whether
there exist polynomial approximation schemesalso for nongeometric and non-
planar covering and packing problems. In general, it would be desired to find
additional strongly NP-complete problems that have polynomial-time approxima-
tion schemes.

REFERENCES
1. BAKER,B. S. Approximation algorithms for NP-complete problems on planar graphs. In Proceed-
ings of the 24th Annual IEEE Symposiumon Foundationsof ComputerScience(Tucson, Ariz.,
Nov. 7-9). IEEE, New York, 1983, pp. 265-273.
2. BERMAN,F. D., LEIGHTON,F. T., AND SNYDER,L. Optimal tile salvage. Unpublished manuscript,
1982.
3. FOWLER,R. J., PATERSON,M. S., AND TANIMOTO,S. L. Optimal packing and covering in the
plane are NP-complete. Inf: Proc.Lett. 12, 3 (June 1981), 133-137.
136 D. S. HOCHBAUM AND W. MAASS

4. CAREY, M. R., AND JOHNSON,D. S. Computers and Intractability: A Guide to the Theory of NP-
Completeness. Fre:eman, San Francisco, 1978.
5. HGCHBAUM,D. S,., AND MAASS, W. Fast approximation schemes for geometric covering and
packing problems in robotics and VLSI. Unpublished manuscript, University of California, Berke-
ley, Calif., 1983.
6. JOHNSON,D. S. The NP-completeness column: An ongoing guide. J. Algorithms 3,2 (June 1982),
182-195.
7. TANIMOTO,S. L., .ANDFOWLER,R. J. Covering image subsets with patches. In Proceedings of the
5th International Conference on Pattern Recognition, 1980, pp. 835-839.

RECEIVEDAUGUST1983; REVISEDAUGUST1984; ACCEPTEDSEPTEMBER

1984

Journal of the Association for Computing Machinery, Vol. 32, No. I, January1985.