DIMACS Series in Discrete Mathematics

and Theoretical Computer Science

Volume 00, 0000
1
The Quadratic Assignment Problem: A Survey
and Recent Developments
PANOS M. PARDALOS, FRANZ RENDL, AND HENRY WOLKOWICZ

Abstract. Quadratic Assignment Problems model many applications in

diverse areas such as operations research, parallel and distributed comput-
ing, and combinatorial data analysis. In this paper we survey some of
the most important techniques, applications, and methods regarding the
quadratic assignment problem. We focus our attention on recent develop-
ments.

1. Introduction
Given a set N = f1; 2; : : : ; ng and n  n matrices F = (fij ) and D = (dkl ),
the quadratic assignment problem (QAP) can be stated as follows:
n X
X n n
X
min
p2
fij dp(i)p(j ) + cip(i);
N
i=1 j =1 i=1
where N is the set of all permutations of N . One of the major applications of
the QAP is in location theory where the matrix F = (fij ) is the ow matrix,
i.e. fij is the ow of materials from facility i to facility j , and D = (dkl) is the
distance matrix, i.e. dkl represents the distance from location k to location l
[62, 67, 137]. The cost of simultaneously assigning facility i to location k and
facility j to location l is fij dkl . The objective is to nd an assignment of all
facilities to all locations (i.e. a permutation p 2 N ), such that the total cost of
the assignment is minimized. Throughout this paper we often refer to the QAP
in the context of this location problem.
In addition to its application in facility location problems, the QAP has been
found useful in such applications as scheduling [88], the backboard wiring prob-
lem in electronics [240], parallel and distributed computing [24], and statistical
data analysis [118]. Other applications may be found in [77, 138, 159].
The term "quadratic" comes from the reformulation of the problem as an
optimization problem with a quadratic objective function. There is a one-to-
one correspondence between N and the set of n  n permutation matrices
1 This paper and a separate bibliography le (bib le) is available by anonymous ftp at
orion.uwaterloo.ca in the directory pub/henry/qap.
1991 Mathematics Subject Classi cation. Primary 90B80, 90C20, 90C35, 90C27; Secondary
65H20, 65K05.
Key words and phrases. Combinatorial optimization, quadratic assignment problem, graph
partitioning, survey, exact algorithms, heuristics, algorithms, test problems, bibliography.
c 0000 American Mathematical Society
0000-0000/00 $1.00 + $.25 per page

1
2 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
X = (xij )nn. The entries of each such matrix must satisfy:
n
X
(1.1) xij = 1; i = 1; : : : ; n;
j =1

n
X
(1.2) xij = 1; j = 1; : : : ; n;
i=1

(1.3) xij 2 f0; 1g; i = 1; : : : ; n; j = 1; : : : ; n;


xij = 1 if facility i is assigned to location j
0 otherwise.
With the above constraints on x, we have the following equivalent formulation
for the quadratic assignment problem, working on the space of permutation
matrices,
n X
X n X
n X
n n
X
(1.4) min aij bkl xik xjl + cij xij :
i=1 j =1 k=1 l=1 i;j =1
The paper is organized as follows. In Section 2 we present the mathemati-
cal tools and techniques that have proven to be useful for QAP. This includes
various formulations of the problem and representations of the feasible set. The
various representations of the feasible set and objective function lead directly to
tractable relaxations. We include optimality conditions and representations of
derivatives for QAP and its relaxations. In Section 3, we present several appli-
cations of QAP, both theoretical and practical. We also include generalizations.
The computational complexity is described in Section 4. A survey of current nu-
merical methods is presented in Section 5. Test problem generation with known
optimal permutation is discussed in Section 6. Concluding remarks are made in
Section 7.
2. Mathematics of QAP
In this section we outline the mathematical tools and techniques that are
useful and interesting for QAP.
2.1. Formulations. Several formulations have been used in the literature to
study the QAP. We outline several of these formulations now. (Please see [100]
for more details and more formulations.)
2.1.1. Koopmans-Beckmann. The QAP was introduced in 1957 by Koopmans
and Beckmann [137] using the formulation presented above in equations (1.1) to
(1.4). This model was formulated to study the assignment of a set of economic
activities to a set of locations.
 THE QUADRATIC ASSIGNMENT PROBLEM 3
2.1.2. Trace. For simplicity, we mainly use the trace formulation in this paper
QAP minX 2 f (X ) = trace (AXB + C )X t ;
where :t denotes transpose,  is the set of permutation matrices, and trace
stands for the sum of the diagonal elements. We assume A and B to be real
symmetric n  n matrices and C 2 <nn. This formulation was introduced in
[65, 66]. It allows for easy manipulation and relaxation of the data and truly
illustrates the n dimensional nature of the problem.
Several elementary properties of the trace can be exploited. For example:
the trace of a symmetric matrix is the sum of the eigenvalues; and the trace of
a product satis es trace MN = trace NM = trace N t M t: Moreover, the trace
provides a valid inner product on the space of real (or complex) m  n matrices
hM; N i = trace M  N;
where
 stands for complex conjugate.
2.1.3. Kronecker Product. The trace formulation of the objective function is
a compact form of representing the quadratic form with the matrix X as the
variable. The Hessian of this quadratic form is the tensor product or Kronecker
product
A B = (aij bkl ) = (aij B ) ;
i.e. the matrix formed from all possible products of elements from A and B .
We can use the notation that vec (X ) 2 <n2 denotes the vector formed from the
columns of the matrix X . Then the objective function is equivalent to
f (X ) = vec (X )t (A B )vec (X ) + vec (C )tvec (X ):
Derivatives and algebraic manipulations can all be done via the Kronecker prod-
uct. For example, the eigenvalues of the Kronecker product are the n2 eigenval-
ues formed from all possible products of the eigenvalues of A and B . However,
using the Kronecker product hides the structure of the problem and increases
the complexity in that we do not take advantage of the hidden fact that we are
really working on an n dimensional problem. For that reason, the Kronecker
product is rarely used and we do not study it further. (See [96] for details on
manipulations and calculus involving Kronecker products.)
2.2. The Feasible Set and Perturbations. The feasible set for QAP con-
sists of all the possible assignments of n objects to n locations. The extreme
points, or vertices, of the bipartite perfect matching polytope (as de ned by the
nonnegative vectors x = (xij ) 2 <n2 satisfying the constraints (1.1) and (1.2))
are the incidence vectors of all possible assignments. Alternatively, in the trace
formulation, the feasible set consists of all the permutation matrices. It is well
4 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
known that the permutation matrices satisfy
(2.1)  = O \ E \ N;
= S \ E \ N;
where: O = fX : X t X = I g is the set of orthogonal matrices; S = fX :
trace X t X = ng; while E = fX : Xu = X t u = ug is the set of all matrices
having row and column sums equal to one; and N = fX : X  0g is the set
of nonnegative matrices. We let D = E \ N denote the set of doubly stochastic
matrices. A well known theorem of Birko [21] states that the convex hull of
the permutation matrices is the set of doubly stochastic matrices, conv  = D:
Thus the set of doubly stochastic matrices corresponds to the bipartite perfect
matching polytope. (For more details on the above, see e.g. [28].)
The properties of the feasible set allow for perturbations of the objective
function without changing the optimal solution of the original problem. These
perturbations do change relaxations for the problem and so are important in
improving bounding techniques. Two standard perturbations, constant row and
column perturbations and diagonal perturbations, are known to have this prop-
erty. Speci cally, suppose that e; f; r; s 2 <n and de ne
A(e; r) = A + eut + uet + diag (r);
B (f; s) = B + fut + uf t + diag (s);
P
C (e; f; r; s) = C + 2Auf t + 2eutB ? 2nef t ? 2 k ek uf t
+diag (A)st + r diag (B )t ? 2est ? 2rf t ? rst ;
where diag changes a vector to a diagonal matrix and, conversely, it changes a
matrix to a vector formed from the diagonal elements. Then
trace (AXB + C )X t = trace (A(e; r)XB (f; s) + C (e; f; r; s))X t; 8X 2 :
(Note that symmetry is preserved by these transformations.)
If we keep the constraint X 2 E , i.e. X has row and column sums equal to
1, then the constant row and column perturbations are redundant and can be
ignored, i.e. only the diagonal perturbations need be used. (See e.g. [102] for de-
tails. The question of which perturbations are needed, under which relaxations,
is discussed in [198].)
2.3. Relaxations. Since QAP is an NP-hard problem, the equivalent ex-
pression of the constraints in (2.1) are very useful. We immediately get repre-
sentations for relaxations. These relaxations remove the combinatorial nature of
the problem and allow for solutions using continuous optimization techniques.
The strategy is to relax the objective function and/or the constraints, in order
to get a tractable problem. These problems do not provide useful approxima-
tions in general. But, we can then nd the best relaxation of a family of these
problems over the above mentioned perturbations. (More details are provided
when we discuss lower bounds in Section 5.2.)
 THE QUADRATIC ASSIGNMENT PROBLEM 5
2.3.1. Linearization. If the quadratic term of QAP vanishes, then we have an
ordinary linear assignment problem which can be solved very eciently. In gen-
eral, the QAP can be relaxed to a (0,1)-linear integer program. This can be done
by introducing new binary variables yijkl = xij xkl : These new variables replace
the occurrence of quadratic terms in the objective function. New constraints are
added to ensure consistency with the original problem. Typically, the convex hull
of the constraint set is used in order to obtain an ordinary linear programming
problem. See e.g. [143, 155, 154, 132, 8, 18, 55, 34, 81, 2, 100].
2.3.2. Quadratic Programming. As stated above in Section 2.2, perturbations
of the objective function can be done without changing the optimum of QAP.
But these perturbations can be very useful in forming relaxations. Since diagonal
perturbations are allowed, we can perturb the diagonal elements of A and B
in order to make the Hessian of the objective function convex. We can then
take the convex hull of the feasible set, i.e. we relax the feasible set to the
doubly stochastic matrices D: This results in a standard quadratic programming
problem that can be solved by well known methods. In fact, we do not need to
make the Hessian positive semide nite on all matrices X , but rather only on the
span of the feasible set, i.e. on the span of the doubly stochastic matrices, or
equivalently, on the span of matrices with row and column sums equal to 1.
However, this relaxation does not fully exploit the structure of the problem,
since it treats the objective function as a quadratic form over <n2 :
2.3.3. Trust Region Subproblems. If we use the second representation in (2.1)
and relax the constraint set to X 2 S \ E ; then we do not have to worry about
convexity of the objective function, i.e. we obtain a tractable problem called a
trust region subproblem. (In [241] it is shown that these problems are really
implicit convex problems, since their dual problems are concave maximization
problems.) However, these problems still do not exploit the structure of QAP.
2.3.4. Parametrization of Permutation and Orthogonal Groups. The vector e
of ones is both a right and left eigenvector corresponding to an eigenvalue of 1,
for every permutation matrix. This fact can be exploited to project the feasible
set of QAP onto the span of the doubly stochastic matrices while not losing the
special trace structure of the objective function.
Let the n  (n ? 1)?matrix V be such that
V te = 0; V tV = In?1:
The columns of V therefore constitute an orthonormal basis of feg? : Further,
let
.
v := keek ; P = [v .. V ] 2 O:
Thus Q := V V t = I ? vvt describes the orthogonal projection on feg? : The
parametrization of the permutation matrices follows. (See [102].)
6 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
Proposition 2.1. Let X be n  n and Y be (n ? 1)  (n ? 1): Suppose that
X and Y satisfy
 
(2.2) X = P 10 Y0 P t:

Then
X 2 E;
X 2 N () V Y V t  ?vvt ;
X 2 On () Y 2 On?1 :
Conversely, if X 2 E ; then there is a Y such that ( 2.2) holds.

Relaxing the constraints to X 2 O or to X 2 O\E removes the combinatorial

nature of the problem. The resulting problem can be split into an eigenvalue
problem for the quadratic part of the objective function and a standard linear
programming problem for the linear part. (This is discussed in detail in the
section on bounds, Section (5.2.2).)
In addition, one can parametrize the orthogonal constraint X 2 O by noting
that it is equivalent to the matrix exponential X = expS for some skew symmetric
matrix S , see e.g. [172, 60]. This results in an unconstrained problem over the
space of skew symmetric matrices.
2.3.5. Semide nite Programming. The orthogonal constraint can be relaxed
to XX t  I , i.e. XX t is negative semide nite. This relaxation is discussed in
detail in an accompanying paper in these proceedings [129].
2.4. Derivatives and Optimality Conditions. QAP is an NP-hard prob-
lem. It is therefore not surprising that verifying optimality is also an NP-hard
problem. In fact, even checking local optimality is a hard problem. (See Section
4.2.) However there are tractable optimality conditions for the relaxations.
2.4.1. Di erentials. We rst present the derivatives of the functions involved
in QAP. Let
k(X ) = XBX t ; g(X ) = X t X ? I; f (X ) = trace AXBX t :
Then the corresponding di erentials in the (matrix) direction h are
dk(X ; h) = XBht + hBX t ;
dg(X ; h) = Xht + htX ;
df (X ; h) = trace A(dk(X ; h)) = trace A(XBht + hBX t ):
 THE QUADRATIC ASSIGNMENT PROBLEM 7
2.4.2. Optimality Conditions. First, consider the relaxation of QAP when
C = 0 and the constraint set consists of the orthogonal matrices. From the
above notation this corresponds to
min f (X ) subject to g(X ) = 0:
The Lagrangian for this problem is
f (X ) + trace Sg(X );
where the Lagrange multiplier S is a symmetric matrix. We can di erentiate the
Lagrangian using the above di erentials. If we set the derivative to 0, we get the
condition that AXB + XS = 0 or X t AXB = ?S: We conclude from S = S t ,
that X t AX and B commute and so are mutually diagonalizable. This yields
the minimum scalar product of the eigenvalues used in the bounds in Theorem
5.1. An improved bound can be obtained by projecting the feasible set onto
the span of the doubly stochastic matrices. This uses a parametrization of the
permutation matrices, see 2.3.4. (More details can be found in [210, 102, 129].)
2.4.3. Global Optimality. If we relax QAP to a quadratic (convex) program,
i.e. we relax the constraint set to the set of doubly stochastic matrices D while
perturbing the objective function to make it convex on the span of D, then the
Karush-Kuhn-Tucker optimality conditions characterize optimality. This is well
known for convex programming problems where a constraint quali cation holds.
The primal relaxed problem is equivalent to the min-max of the Lagrangian

(2.3) Xmin max trace (AXB + C )X t  + t (Xe ? e) + t (X t e ? e);
2N  ; 2<
1 2 n 1 2
while the dual is the max-min problem
 
(2.4)  max min trace (AXB + C )X t + t1 (Xe ? e) + t2 (X t e ? e):
1 ;2 2<n X 2N

The above relaxation provides lower bounds for QAP. Thus, for each pertur-
bation de ned in Section 2.2, with the above convexity assumption on the span
of D, the dual problem provides lower bounds for QAP because we can e ectively
characterize global optimality for it. This is no longer true if we do not make
the convexity assumptions on the objective function.
Statements about global optimality for nonconvex problems, such as QAP
itself, are much harder to make. A characterization for general problems can be
found in [114].
2.4.4. Local Optimality. For the general quadratic programming relaxation,
even in the nonconvex case, i.e. for a general objective function constrained
to D, we still get necessary and sucient local optimality conditions. This is
due to the quadratic nature of the problem, i.e. the second order optimality
conditions are necessary and sucient. However, this is not the case for problems
with nonnegativity constraints. It has been shown in [191] that the problem of
8 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
checking local optimality (and the problem of checking if a local minimum is
strict) in quadratic programming with linear constraints is NP-hard. Pardalos
and Vavasis [192] have also shown that quadratic programming with one negative
eigenvalue (all others zero) remains an NP-hard problem.
2.5. Nonsymmetric QAP. Applications for QAP usually involve symmet-
ric matrices A; B . If one of A or B is symmetric, then we can still get an equiv-
alent symmetric QAP by symmetrizing the other, e.g. replace B by (B + B t )=2:
If both A and B are not symmetric, then we can still symmetrize the quadratic
form by using the Kronecker product, but we then lose the trace structure of the
problem.
However, even if both A and B are not symmetric, we can still obtain mean-
ingful bounds by moving into the space of Hermitian matrices. See [103].
3. Applications, Generalizations and special cases
Applications for QAP are many and varied. Several are mentioned in the
introduction above. We also point out that [33] summarizes recently published
applications of quadratic assignment problem.
We will describe now rst some generalizations of the quadratic assignment
problem, and then discuss some interesting special cases.
3.1. The 3-index Assignment Problem. The three-index (or 3-dimensional)
assignment problem of order n can be stated as a (0,1)-programming problem of
the following form:
P
min P
fcijkxijk : i 2 I; j 2 J; k 2 K g;
s:t: P
fx ijk : j 2 J; k 2 K g = 1; 8i 2 I;
(3.1) P
fx ijk : i 2 I; k 2 K g = 1; 8j 2 J;
fxijk : i 2 I; j 2 J g = 1; 8k 2 K;
xijk 2 f0; 1g; 8i; j; k;
where I , J and K are disjoint index sets with jI j = jJ j = jK j = n. Note that
the number of variables of the three-dimensional assignment problem of order n
is n3.
From the complexity point of view, it has been shown that the three-dimensional
assignment problem is NP-hard [131]. Most ot the proposed algorithms for this
problem are implicit enumeration methods. Some of the proposed algorithms
include those of Vlach [247], Pierskalla [196, 197] and Leue [145]; a primal-
dual algorithm described by Hansen and Kaufman [106]; a branch and bound
algorithm using a Lagrangian dual and subgradient optimization implemented
by Frohlich [83] and discussed by Burkard and Frohlich [38]. Also see Burkard
and Rudolf [42]. More recently, Balas and Saltzman [11] developed a branch
and bound algorithm that also uses facet-de ning inequalities in a Lagrangian
fashion with subgradient optimization.
 THE QUADRATIC ASSIGNMENT PROBLEM 9
Let A be the coecient matrix of the constraint set of (3.1). Then R =
I [ J [ K is the row index set of A. Let S be the column index set of A. Let GA
be the intersection graph of A, i.e., the graph that has a vertex for every column
of A and an edge for every pair of non-orthogonal columns. Let
P = fx 2 Rn : Ax = e; x  0g;
3

where e = (1; :::; 1)t 2 R3n. Then

PI = conv fx 2 f0; 1gn : x 2 P g
3

is the three-index assignment polytope of order n.

Balas and Saltzman [10] started to study the facial structure of PI . They gave
an O(n4) procedure to detect whether there is a clique facet of PI , violated by
a given noninteger point x. In [1], Balas and Qi gave an O(n3 ) procedure to do
this. Since the number of variables of (3.1) is n3 , an O(n3 ) separation algorithm
for a facet class of PI is linear-time and its complexity is best possible. Balas
and Qi [9], Gwan and Qi [99] also gave linear-time separation algorithms for
other two facet classes of PI , identi ed in [10]. More recent results can be found
in [202]].
Other papers on the three-index assignment problem include [43, 68, 80, 81]
and [219].
3.2. The Quadratic Semiassignment Problem. The quadratic semias-
signment problem (QSA) uni es some interesting combinatorial optimization
problems. The general problem has the form:
Pm Pn Pn
min k=1
Pm i=1 j =1 cij xik xjk
(3.2) s:t: k=1 xjk = 1; j = 1; : : : ; n
xjk 2 f0; 1g; 8j; k;
Some special cases of this problem include the clustering problem, the equipar-
tition problem, and the m-coloring problem on graphs [234], [235].
Given n objects and an n  n dissimilarity matrix C = (cij ), the \clustering
problem" is to nd a partition of the objects into m classes (clusters) which
minimizes the sum of the dissimilarities between objects belonging to the same
class.
The \equipartition problem" is the following: Given n objects with weights
wi, i = 1; : : : ; n, nd a partition of the objects into m classes so as to minimize
the variance of the class weights. This problem can be formulated as a quadratic
semiassignment problem with cij = wi wj for all i and j .
The \m-coloring problem" is also a special case of the (QSA) problem. Given
a graph G(V; A), the graph admits a coloration of its vertices with m colors
10 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
(adjacent vertices have di erent colors) if and only if the following problem
Pm P
min Pm k=1 (i;j )2A xik xjk
(3.3) k=1 xjk = 1; j = 1; : : : ; n
s:t: xjk 2 f0; 1g; 8j; k;
has optimal function value zero.
3.3. The Biquadratic Assignment Problem. Recently Burkard, Cela
and Klinz [35] (see the paper in this volume) introduced a fourth order general-
ization of the quadratic assignment problem. Let A = (ai;j;k;l) and B = (bm;p;s;t)
be two arrays of n4 elements. Then the biquadratic assignment problem asks to
minimize
X
ai;j;k;lbm;p;s;t xim xjpxksxlt
i;j;k;l;m;p;s;t
over all permutation matrices X . This problem arises in the eld of VLSI syn-
thesis. In [35] various formulations of this problem are described. Also, lower
bounds and some methods to construct instances with known optimal solution
are presented. There are still many open problems related to this generalized
model of the quadratic assignment problem. In particular it would be interesting
to explore eigenvalue related techniques to this problem.
After having described some relatives of the quadratic assignment problem,
which typically are at least as hard or harder to solve we will now focus on
specially structured quadratic assignment problems which lead to simpli cations
of the problem.
3.4. Special Cases. The quadratic assignment problem can be formulated
very naturally in a graph theoretical context. This formulation was investigated
rst by Christo des and Gerrard [54], and later by Bokhari [23] and Rendl [208].
We review this formulation and present several applications to other optimization
problems on graphs that can be derived from this formulation.
Let G = (V; E ) and G0 = (V 0; E 0) be graphs. G is isomorphic to G0 (G  G0
for short), if there exists an adjacency preserving bijection p : V 7! V 0, i.e.
ij 2 E () p(i)p(j ) 2 E 0 :
We denote by (G; G0) the set of adjacency preserving mappings between G and
G0. Furthermore we denote by M (G; G0) the set of all subgraphs H of G0, which
are isomorphic to G,i.e.
M (G; G0) = fH : H subgraph of G0; H  Gg:
The graph theoretic formulation of quadratic assignment problem was pro-
posed by Christo des and Gerrards as follows.
Let G and G0 be graphs with edge weights a : E 7! <; b : E 0 7! <.
 THE QUADRATIC ASSIGNMENT PROBLEM 11

X
min min
H 2M (G;G ) 2(G;H )
aij b(i)(j ) :
ij 2E
0

(To avoid trivialities one has to assume jV j  jV 0j .) Note that if both

G and G0 are isomorphic to the complete graph Kn , then M (G; G0) = fG0g
and (G; G0) contains all permutations of n elements, leading to the standard
quadratic assignment problem.
The graph theoretic formulation provides many possibilities to generate spe-
cial cases of the standard quadratic assignment problem. The complexity issues
related to these special cases will be addressed in subsection 4. Here we will re-
late other dicult optimization problems on graphs to the quadratic assignment
problem.
It is well known that the traveling salesman problem and the matching prob-
lem can be formulated as a special quadratic assignment problem, see [75]. Per-
haps less known are the connections of the quadratic assignment problem to
the bandwidth problem in graphs. To see this connection we rst introduce the
bandwidth problem (we refer to [201] for a survey on the topic):
Let G be an undirected (and unweighted) graph on n nodes. A permutation 
of n elements is called a labeling of the nodes of G. The bandwidth of a labeling
 is de ned as
max
ij 2E
j(i) ? (j )j:
The bandwidth  of G is the minimum of this number over all labelings. In terms
of matrices, the bandwidth problem asks for a simultaneous permutation of the
rows and columns of the adjacency matrix of G such that all nonzero entries are
as close as possible to the main diagonal.
Suppose the bandwidth of a given graph G is at most k. Let us denote by
Pn;k the graph on n vertices with edges ij whenever ji ? j j  k. Then clearly G
must be isomorphic to some subgraph of Pn;k . Conversely, if the bandwidth of G
is larger than k, then there cannot exist a subgraph of Pn;k which is isomorphic
to G. If we denote the adjacency matrix of G by A and the adjacency matrix of
Pn;k by B , then we conclude: P
The bandwidth of G is at most k if and only if max2 ij aij b(i)(j ) = 2jE j:
Therefore if some upper bound on this quadratic assignment problem has a
value less than 2jE j for a xed value of k, one immediately concludes that the
bandwidth is larger than k. This idea was used in [110] to derive lower bounds
on the bandwidth of graphs. In particular the following simple lower bound on
the bandwidth is proved.
(G)  n2 (L)=n (L) ? 1
Here L denotes the Laplacian matrix of the graph, which is related to A by
L = Diag(Ae) ? A. (Recall that Ae is the vector of row sums of A.)
12 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
A similar approach works also for the problem of bounding the 1-sum of a
graph, which is the same as the bandwidth except that one minimizes
X
j(i) ? (j )j:
ij 2E
over all permutations . Finally the problem of determining whether a vertex
separator of some given size k exists in a graph can be modeled as a quadratic
assignment problem. We refer to further details in [110].
In [166, 209] it is pointed out that the general graph partition problem can
be modeled as a quadratic assignment problem. (See Section 4.3 for details.) It
turns out however, that exploiting the special structure of the partition prob-
lem leads to more powerful results than treating this problem as a quadratic
assignment problem.
4. Complexity Issues and Asymptotic Behavior
From the computational point of view the QAP is one of the most dicult
problems to solve. In this section several aspects regarding the complexity of
the QAP are discussed. Although computational complexity characterizes worst
case instances, it also plays an important role in developing new algorithms for
solving combinatorial optimization problems, analyzing their intrinsic diculty,
and revealing surprising connections among problems and their solutions.
4.1. Computational Complexity. In 1976, Sahni and Gonzalez showed
that the QAP is NP -complete, which implies that nding a polynomial-time
algorithm to solve it is unlikely [221]. In addition, they have also shown that
QAP belongs even to the hardest core of this complexity class, in the sense that
the problem of nding an -approximate solution of QAP remains NP -complete.
Many well known NP -Complete problems, such as the traveling salesman
problem (TSP), the graph partitioning problem (GP), the maximum clique prob-
lem (MCP), can be easily formulated as special cases of the QAP:
 The traveling salesman problem (TSP): The distance matrix corresponds
to the distance matrix of the TSP, the ow matrix corresponds to the
adjacency matrix of a cycle of length n.
 The graph partitioning problem (GP): The distance matrix corresponds
to the adjacency matrix of the GP, the ow matrix corresponds to the
adjacency matrix of two disjoint complete graphs of size n=2 (assuming
n is even).
 The maximum clique problem (MCP): To identify the existence of a
clique of size k, one constructs a distance matrix corresponding to the
adjacency matrix of the graph for the MCP, a ow matrix corresponds
to the adjacency matrix of a clique of size k. The maximum clique can
be found by solving a set of n QAPs, one for each k; 1  k  n.
 THE QUADRATIC ASSIGNMENT PROBLEM 13
There are a few polynomial-time solvable special cases for the QAP. Christo des
and Gerrard [53] investigated the conditions under which the QAP can be solved
in polynomial time. They show that if both matrices A and B are weighted ad-
jacency matrices of a tree, the problem can be solved in a dynamic programming
fashion, in polynomial time. But if only 1 of the 2 matrices is a weighted adja-
cency matrix of a tree, the problem remains to be NP -complete since the TSP
can be reduced to such a form.
Moreover, the following case is solved in O(n log n) running time by Adolphson
and Hu in 1973 [53]: consider the case in which one matrix represents the
weighted adjacency matrix of a tree, while the other one represents the distance
matrix of a grid graph G = (V; E ), where the distances between nodes i and j is
de ned as follows

; if (i; j ) 2 E;
bij = 1length of the short path between i and j , if (i; j ) 62 E:
Other polynomial-time solvable cases include the case in which one of the
matrix is the weighted adjacency matrix of a double star (see Christo des and
Gerrard [53]). When both distance and ow matrices are weighted adjacency
matrices of series-parallel graphs containing no bipartite graph K2;2, then again
the corresponding QAP is solved in polynomial time [208].
4.2. Complexity of Local Search. Next we show that nding a locally
optimal permutation is a dicult problem from the complexity point of view.
A local search algorithm starts with an initial feasible solution and successively
moves to neighboring solutions until no further improvement is possible. To
characterize the complexity of solving combinatorial optimization problems such
as the QAP with local search algorithms, a Polynomial-time Local Search (PLS)
class has been de ned [126] that captures the structure of NP problems at the
level of their feasible solutions and neighborhoods. Similar to NP-completeness,
the concept of PLS-completeness has been de ned to capture the class of the
hardest problems in PLS. For certain NP-complete problems, the corresponding
PLS problems have already been shown to be PLS-complete [126, 223]. In
regard to the complexity of local search, see also [181] and [191].
4.2.1. A K-L Type Local Search Algorithm. Next we describe a new local
search algorithm for the QAP and establish the connection between the new
algorithm and the Kernighan-Lin heuristic algorithm for the (GP).
The local search algorithm for the QAP starts with a random permutation as a
current permutation. For a current permutation p0 , a sequence of permutations,
p1; : : : ; pl , is constructed in a greedy sense. Each of the permutations in the
sequence is obtained from the previous one by swapping (interchanging) two
assignments and has cost lower than the current permutation. A local search is
performed in the sequence, replacing the current permutation by the permutation
with the lowest cost in the sequence (the algorithm stops if the sequence is empty
14 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
for the current permutation). In the description of the local search algorithm
below, instead of using the cost C (pk ) of a permutation pk in the sequence of
permutations corresponding to a current permutation p0, we use the cumulative
gain G(k) of the permutation pk , where G(k) = C (p0 ) ? C (pk ). Hence, the larger
the cumulative gain of a permutation, the lower the cost is.

Algorithm 1: A Local Search Algorithm for the QAP

Input: n, n  n matrices F; D, and a permutation p of size n.
Output: A local optimal permutation p for the QAP.
(i) Set p0 = p and calculate its cost C (p0). Set i = 0; gi = 0, and G(i) = 0,
where gi and G(i) are the step gain and the cumulative gain, respectively.
(ii) i = 1. Initially, select a pair of facilities such that, by exchanging their
locations, a positive step gain is obtained, i.e., g1 = C (p0) ? C (p1) > 0.
If no such pair exists then go to 7, otherwise set G(1) = g1.
(iii) i = i +1. For each pair of facilities not already selected, evaluate the step
gain by exchanging their locations. Then, select the pair with maximum
gain gi = C (pi?1) ? C (pi ). If all facilities have been selected then set
i = i ? 1 and go to 5.
(iv) Compute the cumulative gain, G(i) = kk==1i gk . If G(i) > 0; then go to
P

3.
(v) Select k, such that G(k) is maximum for 0  k  i.
(vi) If k > 0 then set p0 = pk and go to 2.
(vii) We have reached a local optimum for the QAP. Set p = p0 and output
p and C (p).

Now let us review the (KL) heuristic algorithm for the GP for an undirected
graph G(V; E ) (assuming jV j = 2n) with edge weights w(e), e 2 E . For conve-
nience, a partition of the set V always means a partition into two sets (A; B )
with jAj = jB j = jV j=2 in the rest of the paper. Then, the problem GP is
to nd a partition (A; B ) of the set V with the minimum cost C (A; B ), which
is de ned to be the sum of the weights of all edges between A and B . As
one of the most successful heuristic algorithms for the GP, the Kernighan-Lin
heuristic starts with a random partition of the set V . A sequence of partitions,
(A1 ; B1); : : : ; (Al ; Bl ), is constructed for a current partition (A0 ; B0 ) in a greedy
sense. Each partition (Ak ; Bk ); 1  k  l, in the sequence is obtained from the
previous one (Ak?1; Bk?1) by swapping one vertex in Ak?1 with one vertex in
Bk?1 and has cost lower than the current partition. A local search is performed
in the set of partitions of this sequence, replacing the current partition by the
partition with the lowest cost in the sequence (the algorithm stops if the sequence
is empty for the current partition). Similar to the description of Algorithm 1,
we use the cumulative gain G(k) for a partition pk .
 THE QUADRATIC ASSIGNMENT PROBLEM 15

Algorithm 2: Kernighan-Lin heuristic for the GP

Input: n; G = (V; E ) with jV j = 2n; W = (wij ), and a partition (A; B ) of V .
Output: A locally optimal partition (A; B ) of V
(i) Set A0 = A and B0 = B , obtain its cost C (A0 ; B0). Set i = 0; gi = 0,
and G(i) = 0, where gi and G(i) are step gain and cumulative gain,
respectively.
(ii) i = 1. Initially, select a pair of vertices a1 2 A0 and b1 2 B0 such that,
by swapping them, the resulting partition (A1 ; B1) produces a positive
step gain g1, i.e., g1 = C (A0; B0 ) ? C (A1 ; B1) > 0. If such a pair does
not exist then go to 7, otherwise set G(1) = g1.
(iii) i = i +1. Among the vertices not selected so far, choose a pair ai 2 Ai?1
and bi 2 Bi?1 and swap them to obtain Ai and Bi with maximum step
gain gi = C (Ai?1; Bi?1)?C (Ai ; Bi). If all the vertices have been selected
then set i = i ? 1 and go to 5.
(iv) Compute the cumulative gain G(i) = kk==1i gk . If G(i) > 0; then go to
P

3.
(v) Choose k, such that G(k) is maximum for 0  k  i.
(vi) If k > 0 then set A0 = Ak and B0 = Bk and go to 2.
(vii) We have reached a local optimum for the GP; set A = A0 and B = B0 .
Output A; B and C (A; B ).

Comparing the above algorithm with the local search algorithm for the QAP,
one can easily see the similarity between them. Instead of working with partitions
in the GP, we work with permutations in the QAP. The reduction from the GP to
the QAP in the next section reveals why the adaptation of (KL) algorithm to the
QAP can be e ective. Furthermore, extensive computational results in section
3 indicate that the proposed local search algorithm (Algorithm 1) performs very
well.

4.3. PLS-completeness and the QAP. For many combinatorial optimiza-

tion problems, local search gives rise to some of the most successful heuristics.
A classical example in this regard is the Linear Programming Problem for which
the Simplex method can be viewed as a local search algorithm, in which a local
search step is to go from the current basis to an adjacent basis which di ers from
the current one by one column vector. Based on the pivoting rule, worst-case
examples can be constructed that force the Simplex method to take exponential
time. Whether there can be a pivoting rule under which the Simplex method
takes only polynomial time is a major open question.
In order to characterize the complexity of such local search algorithms, a new
complexity class, the Polynomial-time Local Search class, was introduced and
16 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
studied in [126]. A problem P is in PLS if, for each instance x 2 I (the set
of all instances), we have a set of feasible solutions F (x) such that it is easy to
decide whether s 2 F (x) for any solution s. Then, given x 2 I , we can produce
a feasible solution s 2 F (x) in polynomial time. Next, given x 2 I and s 2 F (x)
we can compute the cost C (s; x) of s in polynomial time. In addition, every
solution s 2 F (x) has a set of neighboring solutions N (s; x). Finally, given x 2 I
and s 2 F (x), we can test in polynomial time whether s is locally optimal, and if
not, produce a solution belonging to N (s; x) with a better cost value (A solution
s is locally optimal if it does not have a strictly better neighbor).
More formally, a local search problem P in PLS is de ned as follows: Given
an input x, nd a locally optimal solution s 2 F (x). For the problem P , the
following three polynomial time algorithms should also exist.
(i) Algorithm A, on input x 2 I , computes an initial feasible solution s0 2
F (x).
(ii) Algorithm B, on input x 2 I and s 2 F (x) computes C (s; x).
(iii) Algorithm C, on input x 2 I and s 2 F (x), either determines that s is
locally optimal or nds a better solution in N (s; x).
A problem P 2 PLS is PLS-reducible to another problem Q 2 PLS, if there
are polynomial time computable functions f and g, such that f maps an instance
x of P to an instance f (x) of Q and for any locally optimal solution s for f (x),
g(s; x) produces a locally optimal solution for x. A problem P in PLS is PLS-
complete, if every other problem in PLS is PLS-reducible to P .
An example of a PLS-complete problem is the GP with the (KL) neighborhood
structure de ned in Algorithm 2 [126]. The (KL) neighborhood of a partition
for the GP can be de ned as follows. A swap of a partition (A; B ) is a par-
tition (A ; B ), where A and A have a symmetric di erence of 2, i.e., (A ; B )
0 0 0 0 0

is obtained from (A; B ) by swapping one element of A with one element of B .

(A ; B ) is a greedy swap if C (A; B ) ? C (A ; B ) is maximized over all swaps of
0 0 0 0

(A; B ). If in fact (A ; B ) is the lexicographically smallest over all greedy swaps,

0 0

we say that (A ; B ) is the lexicographic greed swap of (A; B ). Let (Ai ; Bi ) be a

0 0

sequence of partitions, each of which is a swap of the one preceding it, starting
from (A0 ; B0). We call it monotonic, if the di erences of Ai ? A0 and Bi ? B0
are monotonically increasing (that is, no vertex is switched back to its original
set (A0; B0 )). Finally, we say that a partition (A ; B ) is a neighbor of (A; B ) if
0 0

it occurs in the unique maximal monotonic sequence of lexicographically greedy

swaps starting with (A; B ). Note that such a sequence will consist of jV j=2 + 1
partitions, with the last one equal to (B; A). Thus, each partition has jV j=2
neighbors. The algorithm performs local search over this neighborhood struc-
ture, replacing the current partition by the partition with the lowest cost in the
neighborhood.
In the remaining part of this section, we show that the QAP with the neigh-
borhood structure de ned in Algorithm 1 is PLS-complete by reduction from
 THE QUADRATIC ASSIGNMENT PROBLEM 17
the GP with the (KL) neighborhood structure. First, we show that the local
search problem for the QAP is in PLS. Since the set of feasible solutions of the
QAP is the set of permutations, an initial feasible solution can be produced in
linear time. Computing the cost of a permutation for the QAP can be done
in polynomial time. The neighborhood structure de ned for the QAP in Algo-
rithm 1 is quite similar to the (KL) neighborhood structure for the GP. For a
given permutation for the QAP, there are bn=2c neighbors. We can determine
in polynomial time if the permutation is locally optimal, and if not, produce
a better permutation among the bn=2c neighboring permutations. Hence, with
this neighborhood structure, nding a local optimum for the QAP is in PLS.
To prove PLS-completeness, we show that the GP is PLS-reducible to the
QAP. Given an instance of the GP of size 2n, we can create an instance of the
QAP with the same size in polynomial time. Furthermore, for each local optimal
permutation of the QAP, there is a natural local optimal partition for the cor-
responding GP. More speci cally, suppose for the GP, the graph G = (V; E ) has
edge weights w(e) and vertex set V with jV j = 2n. We construct, in polynomial
time, an instance of the QAP with 2n  2n matrices F = (fij ) and D = (dkl )
de ned below:

fij = w(i; j ) if (i; j ) 2 E ; otherwise fij = 0;

dkl = 0 if k; l 2 A or k; l 2 B ; otherwise dlk = 1;

where A = f1; 2; : : : ; ng; B = fn + 1; n + 2; : : : ; 2ng:

This reduction de nes a one-one correspondence between a permutation pk of
the QAP with a partition (Ak ; Bk ) of the vertex set V of the corresponding GP.
The set of facilities allocated to locations 1 to n in pk constitutes the set Ak .
The set of facilities allocated to locations n +1 to 2n in pk constitutes the set Bk .
The cost of pk for the QAP is exactly twice the cost of the partition (Ak ; Bk ) for
the GP. Let the partition corresponding to a permutation p0 be (A0 ; B0 ), then
a permutation pk is a neighboring permutation of p0 if, and only if, (Ak ; Bk ) is
a neighboring partitions of (A0 ; B0 ). Hence, for any local optimal permutation
of the QAP, the corresponding partition is a local optimal partition for the GP
and can be recovered in polynomial time. By de nition, the local search problem
for the QAP with the neighborhood structure de ned in Algorithm 1 is PLS-
complete.
We should also mention that, at present, there are no known local criteria
in deciding how good a local optimal solution is, in relation to the global opti-
mum. From the complexity point of view, it can be shown that, if there exists a
polynomial time algorithm for checking whether a given permutation is globally
optimal, then P = NP [177].
18 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
4.4. Asymptotic Behavior. A nice feature of the QAP is that the relative
di erence between the worst and optimal solutions becomes arbitrarily small
with a probability tending to 1 as the problem size tends to in nity. Burkard
and Finke discovered this behavior for the QAP in the plane, i.e., the distance
matrix B corresponds to Euclidean distances in the plane. Later they showed
that this behavior holds also for the QAP in general in 1985. The result can be
stated in the following theorem.
Theorem 4.1. For i; j; k; l 2 f1;


ng, let cijkl be identically distributed ran-
dom variables in [0; 1] with expected value E and variance 2 > 0. For every
xed permutation p 2 , let cip(i)jp(j ) be independently distributed. For given
 > 0 and 0 < 0  2 and 0 < (E + 0 )=(E ? 0)  1 + ,
+
P ( FC? < 1 + )  1 ? 2n!e?0n ;
2
F
C
where 0 = 2((0 )=(0 + 22 ))2 ; limn!1 n!e?0n2 = 0, and FC+ and FC? are the
maximum and minimum objective function values for the QAP with cost matrix
C.
Several other researchers, including Frenk, van Houweninge, and Rinnooy Kan
[78], and Rhee [211, 212] improved the order of convergence and showed that
the convergence holds almost everywhere.
5. Methods of Solution
5.1. Exact Algorithms. In this section, we describe di erent methods used
to nd an optimal solution of QAP. The methods include dynamic programming,
cutting plane, and branch-and-bound techniques.
Among these methods, branch-and-bound is the most successful one, on which
this section is focused the most. Currently, problems of size greater than 15 are
generally dicult to solve. This is due to the inherent diculty of the QAP,
characterized by the lack of sharp lower bounding techniques for moderate and
large size problems. For this reason, a separate section is devoted to lower
bounding techniques for the QAP.
Cutting plane methods for the QAP were introduced by Bazaraa and Sherali
[19]. Although the computational experience was not satisfactory, such methods
can be used to nd good suboptimal solutions, see e.g., Burkard and Bonniger
[34]. In Bazaraa and Sherali [20], cutting plane procedures were investigated
for solving the concave quadratic minimization formulation of the QAP. Sev-
eral heuristics derived from the cutting plane procedures produce good quality
solutions in early stage of the search procedure.
Christo des and Benavent [52] used a special dynamic programming approach
for the special case of the QAP in which the ow matrix is the weighted adjacency
matrix of a tree. Problems of sizes up to 30 were solved.
 THE QUADRATIC ASSIGNMENT PROBLEM 19
Branch-and-bound is a general technique for solving combinatorial optimiza-
tion problems. To solve the QAP with branch-and-bound, currently there are 3
types of algorithms: single assignment algorithms, pair assignment algorithms,
and the relative positioning algorithm. They all start with the empty permuta-
tion as the initial partial permutation; during the execution of the algorithms,
the partial permutation is extended to a full permutation. Single assignment
algorithms date back to Gilmore [89] and were extended to the general QAP by
Lawler [143]. In this approach, as described in the work by Gilmore and Lawler,
a facility is assigned to a location at each node of the branch-and-bound search
tree. Some of the earliest branch and bound algorithms for solving QAPs are
described in [36], [66], [180] and [218].
Pair assignment algorithms were developed by Gavett and Plyter [86], Land
[141], and Nugent et al. [171], etc. At each node of the branch-and-bound
search tree, a xed pair of facilities is allocated to a pair of locations. The last
algorithm, the relative positioning algorithm, was developed by Mirchandani
and Obata [160]. In their approach, the levels of the branch-and-bound search
tree do not correspond to the assignments of facilities to locations. The partial
permutations at each level are determined in terms of distances between facilities,
i.e., their relative positions.
Numerical experiences indicate that among the 3 types of branch-and-bound
algorithms the single assignment algorithms are the best. The pair assignment
algorithms were shown to be not computationally ecient. The authors of the
relative positioning algorithm claimed favorable behavior of the algorithm for
problems with sparse matrices.
5.2. Lower Bounds. Lower bounds are keys to the success of a branch-and-
bound type algorithm in combinatorial optimization. The ideal lower bounds
should be sharp and should be fast to compute. For the QAP, there are roughly
3 categories of lower bounds. The rst category includes the classical Gilmore-
Lawler bound (GLB) and related bounds [89, 143]. The second category in-
cludes the eigenvalue based bounds [74, 102, 101, 210, 103]. The rest of the
bounds are mostly based on reformulations of the QAP and generally involves
solving a number of linear assignment problems [5, 47, 54, 82]. It is generally
acknowledged that the eigenvalue based bounds are the best but also the most
expensive to compute. In the following, a brief discussion of the 3 categories of
lower bounds is given.
5.2.1. Gilmore-Lawler Bound (GLB) and Related Bounds. The GLB is com-
puted by using the minimal vector product and the maximal vector product,
denoted hx; yi? and hx; yi+ , de ned below
hx; yi? = Pmin
2hx; Pyi; hx; yi+ = max
P 2
hx; Pyi;
where the set  denotes the set of all permutations of N and x; y 2 Rn . In
fact, hx; yi? can be computed as the inner product of x+ and y? , where x+ is
20 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
obtained by ordering the components of x ascendingly and y? is obtained by
ordering the components of y descendingly. hx; yi+ can be computed similarly.
Let ai ; bi; i = 1; :::; n, represent the row vectors of matrices A; B , respectively.
Let a^i be the vector consisting of the (n ? 1) components of ai , not including aii .
Let ^bi be the vector consisting of the (n ? 1) components of bi , not including bii .
De ne a matrix L = (lij ) as follows
lij = aii bjj + ha^i ; ^bj i? ; i; j = 1; :::; n:
Then GLB(A,B), the GLB for QAP(A,B), is de ned to be the solution to the
linear assignment problem (LAP) with cost matrix L, i.e.
n
X
GLB (A; B ) = min
p2
lip(i) :
i=1
5.2.2. Eigenvalue Based Bounds. Bounds based on eigenvalues of the ow and
distance matrices A and B have been proposed in a series of papers by Finke
et al. [74], Hadley et al. [102, 101], and Rendl and Wolkowicz [210]. These
bounds, denoted by EVB, are based on the trace formulation of the QAP, see
(2.1.2).
A lower bound for the quadratic part of QAP, based on eigenvalues of A and
B , is given by relaxing the constraint set of permutation matrices to O. This
results in the following theorem (see [70, 74]).
Theorem 5.1. Let A and B be symmetric matrices, and 1  2 :::  n be
the eigenvalues of A, and 1  2 :::  n be the eigenvalues of B . For any
p 2 , n n X n n
X X X
i n?i+1  aij bp(i)p(j )  i i :
i=1 i=1 j =1 i=1
The linear part of QAP is bounded exactly by solving a linear sum assignment
problem, denoted LSAP(C). We get the following bound for QAP
n
X
EV B fA; B g = i n?i+1 + LSAP (C ):
i=1
Certain reductions of the original matrices have to be performed before using
the eigenvalues to obtain lower bounds for the QAP. These involve diagonal,
or constant row and column, perturbations of the matrices. Several such lower
bounds, EVB1, EVB2, EVB3, and IVB, were developed in [74, 101].
The strengthened relaxation of the constraint set of permutation matrices to
O \ E was done in [101]. This relaxation proved to be particularly ecient and
made the constant row and column reductions redundant. Rendl and Wolkow-
icz [210] recently proposed a new lower bound (MEVB) based on eigenvalue
decomposition in conjunction with a steepest ascent algorithm. The bound is
obtained iteratively, with each iteration taking O(n3 ) running time. This latter
technique attempted to combine the bounds for the linear and quadratic parts.
 THE QUADRATIC ASSIGNMENT PROBLEM 21
(A further attempt to avoid taking the sum of two minima obtained by treating
the quadratic and linear parts separately is given in the paper by Karisch, Rendl,
and Wolkowicz in these proceedings.)
5.2.3. Reformulation Based Bounds. There are also other types of lower bounds
for the QAP. Assad and Xu [5] proposed a bound (AX) for a class of quadratic
0-1 programs, including the QAP. The bound is obtained iteratively. In each
iteration, n2 + 1 assignment problems of size n are solved. Hence the run-
ning time to compute the bound is O(kn5) where k is the number of iteration.
Christo des and Gerrard [54] proposed a lower bound (CG) by solving O(n4)
linear assignment problems corresponding to pairs of assignments, resulting in a
O(n7) procedure. Frieze and Yadegar [82] obtained 2 lower bounds by solving
the Lagrangian relaxation of a related linear integer formulation of the QAP.
The bounds are denoted by FY1 and FY2 respectively. Finally, Carraresi and
Malucelli [47] proposed a new lower bound (CM) for the QAP through an itera-
tive process. In each iteration, at most O(n2) linear assignment problems related
with an equivalent reformulation of the QAP are solved. Hence the procedure
has a time complexity of O(kn5 ) where k is the number of iterations used.
5.2.4. A New Class of Lower Bounds. A class of lower bounds based on opti-
mal reduction schemes for the QAP was proposed in Li, Pardalos, Ramakrishnan
and Resende [149]. For a given QAP(A,B), consider a partition of A into two
matrices A1 = (a(1) (2)
ij ) and A2 = (aij ) such that A = A1 + A2 and a partition of
B into two matrices B1 = (b(1) (2)
ij ) and B2 = (bij ) such that B = B1 + B2 . For
each pair (i; j ); i; j = 1; :::; n, consider the following minimization problem
n
X n
X n
X n
X
(5.1) min a(1) (1)
ik bjp(k) + a(2)
ki bp(k)j + akib(2)
p(k)j ? a(2) (2)
ki bp(k)j
k=1 k=1 k=1 k=1
where p 2  and p(i) = j:
De ne a n  n matrix L = (lij ) where lij is the optimal objective function value
of (5.1). The following theorem de nes a new lower bound [149, Theorem 4.1].
Theorem 5.2. Let the matrix L be de ned as above. Then the solution of
the linear assignment problem with cost matrix L is a lower bound for the corre-
sponding QAP.
The classical Gilmore-Lawler bound is a special case in which both matrices
A and B are not partitioned. Di erent ways of partitioning the matrices A and
B (we also refer to this as reduction) yield di erent lower bounds. The common
reduction techniques used in the literature choose A2 and B2 with constant
column sums (which we call constant columns). We refer to such techniques as
constant column reductions.
Let M = (mij ) be a matrix in Rnn. We treat a row vector mi ; 1  i  n,
of M as a 1  n matrix and a column vector mtj ; 1  j  n as a n  1 matrix.
22 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
For convenience of discussion, we use the following notation for average (M ),
variance V (M ), and total variance T (M; ) of M :
n X n n X n
(M ) = n12 mij ; V (M ) = n12
X X
( (M ) ? mij )2;
i=1 j =1 i=1 j =1
m
X
T (M; ) =  V (ai ) + (1 ? )V (M ); for 0    1:
i=1
Note that the statistical total variance used here is the convex combination of
two variances, (a) average row variance, (b) variance of the entire matrix, i.e.
 Xn
T = n V (mi ) + (1 ? )V (M )
i=1
n n X n

= n ( (mi ) ? mij )2 + n2
X X
( (M ) ? mij )2
i=1 i=1 j =1
1
= n2 T (M; ):
In our reduction scheme, we considered the partition A = A1 + A2 , where A1 =
A +
and A2 = ?
, such that the variances of A1 and A2 , the sum of variances
of the rows of A1, and the sum of variances of the rows of A2 are minimized.
This minimization problem has been formulated as follows:
min T (A +
; ) + (1 ? )T (?
t ; )
such that
2 Rnn
where 0    1.
Motivated by the observation that for QAP(A,B) the smaller the variances
of A and B are the tighter the GLB is, the following reduction schemes were
proposed in Li, Pardalos, Ramakrishnan and Resende [149]:
R-1) a(1) (2)
ij = aij ? (ann ? aij ) and aij = (ann ? aij ); i; j = 1; :::; n:
R-2) a(1) t t (2) t t
ij = aij ? ( (an ) ? (aj )) and aij = ( (an ) ? (aj )); i; j = 1; :::; n:
One new lower bound proposed in [149] is to use the reduction scheme R-
1. We denote this lower bound by LB1(). The other new lower bound that
we propose is to use the reduction scheme R-2. This lower bound is denoted
LB2(). Both new lower bounds are dependent on the parameter . Note that,
LB1(0.0) = GLB (A; B ) and LB1(1.0) = GLB (At ; B t ).
For LB1(), we found in our computational experiments that  = 0:5 is a
good choice. For LB2(), we used  = 1:0. The latter was expected since the
column variance of the matrix
is already zero when computing LB2().
The new lower bounds can be computed quite eciently. Computing the
matrix
to partition matrices A and B takes only O(n2) time. By presorting
the rows of the ow and distance matrices A and B , one can compute lij ; i; j =
 THE QUADRATIC ASSIGNMENT PROBLEM 23
1; :::; n, in O(n3) [149]. Hence the total running time is O(n3 ), which is the same
as that for computing GLB. Furthermore, the constant factor is small.
Recently, Jansen [122] has derived an analytical closed form solution to (5.2{
5.2). That solution is given by
ij =  11??  (a ) + (1 ? ) + 2 (1 ? ) ? 2 2 (1 ? ) (A)
i (1 ? )(1 ?  + )
 (1 ?  )
? 1 ?  +  (atj ) ? aij :
Note that, as  ! 1, this partitioning scheme approaches the constant col-
umn reduction partitioning. Experimentally, we have observed that the column
reduction partitioning scheme is more e ective and is easier to implement. For
an ecient implementation of the lower bound and computational results of a
branch and bound algorithm that uses the new bound see [148].
5.3. Suboptimal Algorithms. As exact algorithms can only solve small
size instances of the QAP, and nding an -approximate solutions of the QAP
remains NP -complete, heuristics with good performance in both solution time
and solution quality are highly desirable. Research in this direction abounds
in literature. Basically, there are 5 types of heuristics including construction
methods which start from an empty permutation p and expand p to a subop-
timal permutation according to certain criteria, limited enumeration methods
which perform partial enumeration with the expectation that good solutions
are generally found in early stages of enumeration, improvement methods which
start from a permutation p and try to get an improved permutation by using
some techniques, simulation approaches and genetic algorithms which belong to
stochastic search techniques.
5.3.1. Construction Methods. Construction methods, as the name suggests,
construct a suboptimal permutation step by step. The basic idea was rst in-
troduced by Gilmore [89]. The general scheme for a construction method is as
follows. It starts with a partial permutation p which is empty. Then p is ex-
panded by repetitively selecting a pair of assignment (i; j ) such that i 62 M and
j 62 p(M ) according to certain heuristic, where M is the index set containing
the indices of p for which the corresponding assignments are done and p(M ) is
the set fp(i) j i 2 M g. This process is repeated until p becomes a complete
permutation.
One of the oldest heuristics used is the CRAFT (Computerized Relative Al-
location of Facilities Technique) [4, 231, 30]. This is a well-known heuristic for
designing the layout of facilities that has been in use for over 25 years. Given
a set of departments, locations, a matrix of ows between departments, and
a matrix of costs to transport one item between two departments a unit dis-
tance, CRAFT iteratively improves an initial, user-supplied, layout by a series
of department exchanges. At each step CRAFT considers either all possible 2-
24 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
way, 3-way, or both 2-way and 3-way exchanges. It chooses the exchange that
provides the most improvement in minimizing total cost, and then repeats the
process until no improving exchange can be found.
5.3.2. Limited Enumeration Methods. The argument for such type of methods
is that a good suboptimal solution may be produced early in an enumerative
search procedure, while nding the optimal solution takes much longer time. For
example, it was observed in an experiment that one of the Nugent test problems,
of size n equal to 15, the optimal solution was found after 23.48 seconds on a
CDC Cyber 76, yet it took more than an hour to prove the optimality of the
solution.
There are several ways to limit enumeration of the set of all possible permuta-
tions. One simple way is to put a time limit on the search procedure. Enumera-
tion stops when a predetermined time limit is reached or there is no improvement
within a certain time limit. Another way is to decrease the requirements for op-
timality. For example, whenever an improvement is not obtained after a given
time period, the upper bound is decreased by a certain speci ed percentage, re-
sulting in deeper cuts in the enumeration tree. Although it is possible that the
optimal solution may be cut o , the enumeration process is speeded up. Fur-
thermore, one can estimate that the optimal value di ers from the suboptimal
one by not more than some speci ed percentage.
5.3.3. Improvement Methods. The majority of the heuristic solution methods
for the QAP falls into the category of improvement methods. An improvement
method starts with a solution and tries to move to a better solution while im-
proving the solution quality. Local search algorithms and tabu search belong to
such type of methods. There is rich literature on local search algorithms for sub-
optimal solutions for the QAP. Local search algorithms are iterative algorithms.
At each iteration, a local search algorithm generates the next solution by nding
a better solution in the neighborhood of the current solution. The algorithm
terminates when there is no better solution in the neighborhood of the current
solution.
Tabu search was introduced by Glover [90, 91] as a technique to overcome lo-
cal optimality in combinatorial search. The underlying idea is to limit the search
directions for each search step to obtain good quality solutions in an e ective
way. This approach has been applied successfully to a number of combinato-
rial optimization problems including the TSP. Adaptations of tabu search to the
QAP have been studied by Skorin-Kapov [238] and Taillard [242]. The basic
idea is as follows.
To improve a given initial permutation tabu search seeks, among the set of
permutations obtained by a pair exchange of assignments, a permutation with
the best heuristic evaluation. In the simplest case, such an evaluation dictates
the choice of a permutation which give the best objective function value. Every
choice of a neighboring permutation represents an exchange of a pair of facilities.
 THE QUADRATIC ASSIGNMENT PROBLEM 25
A tabu list with a speci ed size is maintained to forbid certain choices of neigh-
boring permutations, i.e., to forbid certain exchanges between facilities. The
tabu list is essentially a list of pair exchanges occurring in the search process.
As certain valuable exchanges of facilities may be prohibited by the tabu list, an
aspiration function is introduced to allow exchanges prohibited by the tabu list
to be chosen if they are judged to be valuable.
5.3.4. Simulation Approaches. An interesting analogy between problems in
combinatorial optimization and statistical mechanics has been developed and
has proven useful in solving some traditional optimization problems, such as
computer design, partitioning, component placement, wiring, and the traveling
salesman problem. The analogy has resulted in a methodology, termed simulated
annealing, which is used to overcome local optimality (see Kirkpatrick, Gelatti,
and Vecchi [135]).
The term \annealing" refers to the process of a thermal system by rst melt-
ing at high temperatures and then lowering the temperatures slowly based on
an annealing schedule. The process is continued until the vicinity of the solidi -
cation temperature is reached, where the system is allowed to reach the \ground
state" (the lowest energy state of the system). Simulated annealing is a Monte
Carlo approach to simulate the behavior of this system to achieve thermal equi-
librium at a given temperature in a given annealing schedule. This analogy has
been applied in solving combinatorial optimization problems. According to the
above authors:
Iterative improvement, commonly applied to such problems, is much like the
microscopic rearrangement process modeled by statistical mechanics, with the
cost function playing the role of energy. However, accepting only rearrangements
that lower the cost function of the system is like extremely rapid quenching high
temperatures to T = 0. So, it should not be surprising that resulting solutions
are usually metastable. The Metropolis procedure from statistical mechanics
provides a generalization of iterative improvement in which controlled uphill
steps can also be incorporated in the search for a better solution.
Simulated Annealing is applied to the QAP [253] as follows:
 Given any feasible solution (a permutation of locations in relation to
the facilities), randomly select two facilities, make a pair exchange and
evaluate the consequent change (f ) in the total cost (f )
 Repeat the above step as long as f < 0. Otherwise, select a ran-
dom variable x from a uniform distribution U (0; 1). If x < P (f ) =
EXP (?f=ti ) (where P represents the probability obtained from the
exponential distribution (EXP)), then accept the pair exchange and re-
peat the process. Here ti represents the annealing schedule temperature
at stage i where t1 > t2 > : : : > tr represents the annealing schedule.
For example ti = 10  (0:9)(i?1).
 The system remains at stage i until a predetermined number of pair
26 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
exchanges have been considered before going to the next stage.
 If all the temperatures in the annealing schedule have been used, i.e. if
i > r, then stop.
In the search process for an optimal permutation for the QAP, the simulated
annealing searches through the 2-change neighborhood of a present permutation.
It uses Monte Carlo sampling to occasionally accept solutions which increase
rather than decrease the objective function value. Such a choice of neighboring
permutation is counter to the normal steepest descent strategy. However, it is ar-
gued in the analogy that by taking such controlled ascent steps, the optimization
algorithm needs not get stuck on poor solutions.
Burkard and Rendl [41] applied simulated annealing to the QAP and reported
favorable computational results. Wilhelm and Ward [253] further investigated
the procedure. Essential to the success of the adaptation of simulated annealing
to the QAP is the annealing schedule as discussed in their work [41, 253].
In the paper [183], computational results with four heuristics, the CRAFT,
simulating annealing, tabu search, and a local search based on graph partitioning
are reported.
5.3.5. Genetic Algorithms. Genetic algorithms, like simulated annealing, form
another type of stochastic search technique. While simulated annealing is based
on thermodynamic process, genetic algorithms are based on the mechanics of
natural selection and natural adaptation. A genetic algorithm maintains a pop-
ulation consisting of a subset of individuals (solutions). Through means of biased
selection and genetic operations, the algorithm replaces a population with a new
population of individuals with better tness values on the average.
Genetic algorithms have been developed by John Holland et al. at the Uni-
versity of Michigan in 1975. However, genetic algorithms did not have a major
in uence at that time. With the advent of parallel computers, there has been
increasing interest in genetic algorithms since they are inherently parallel. A
number of researchers have tried to apply genetic algorithms to solve combina-
torial optimization problems, such as the the graph partitioning problem and
the traveling salesman problem [161].
5.4. Greedy Randomized Adaptive Search Procedures (GRASP).
GRASP is an iterative randomized sampling technique in which each iteration
provides an approximate solution to the problem at hand. The incumbent solu-
tion over all GRASP iterations is kept as the nal result. There are two phases
within each GRASP iteration: the rst constructs an initial solution via an adap-
tive randomized greedy function; the second applies a local search technique to
the constructed solution in hope of nding an improvement. A comprehensive
survey of GRASP can be found in [73].
In [71] GRASP has been applied to a quadratic assignment problem that
models the positioning of intermodal highway trailers on railcars. The GRASP is
 THE QUADRATIC ASSIGNMENT PROBLEM 27
incorporated within a branch and bound algorithm to compute optimal solutions.
In Li, Pardalos et al (see the paper in this volume), the GRASP has been applied
to solve the general QAP. The GRASP was tested on 88 instances of QAPs (most
of which are from QAPLIB) and found the best known solution of almost all the
instances, and improved on the best known solution in a few cases.
6. Test Problem Generation
In this section we consider the problem of generating QAPs with a known
optimal permutation. The problems will be of the form: Given a set N =
f1; 2; : : : ; ng and two (n  n) matrices F = (fij ) and D = (dkl ), nd a permu-
tation p of the set N that minimizes:
XX
QFD (p) = fij dp(i)p(j ):
i j
Test problem generators are very important in computational optimization.
The eciency of an algorithm for a given problem is determined by several
criteria including the accuracy of the solution, the speed of the algorithm, and the
e ectiveness of the algorithm with respect to di erent problem classes. However,
for many dicult problems, existing theory cannot itself provide measurement
for these criteria. Hence, empirical computational experimentation is necessary.
Evaluation and test of an algorithm can be done by using test problems with
a known optimal solution. Test problems also provide a standard platform on
which di erent algorithms for the same problem can be compared. For general
references on this subject, see e.g., Pardalos [178, 179], Pardalos and Rosen
[190], and Floudas and Pardalos [76].
6.1. Palubetskis' Generator for QAPs with a Known Solution. Next,
we discuss the generation of test problems for the quadratic assignment problem.
One of the rst methods for constructing test problems with a known optimal
permutation was proposed by Palubetskis [175]. Assume that the distance ma-
trix is taken from a grid graph.
Input: w, a value to initialize the F matrix, and z < w, to obtain random values
between [0; z ].
Output: Matrices F and D and an optimal permutation p .
(i) Construct the matrix D = (dij ), of which the elements are the distances
between the knots of the two dimensional grid r  s, where rs = n, using
rectilinear distances. If (i; j ) are neighboring knots, then dij = 1.
(ii) Set F = (fij ) where fij = w (an input parameter to the algorithm).
Compute gij = 2 ? dij .
(iii) While for any i; j = 1; : : : ; n such that gij  0.
(iv) Choose the pair l; m, such that dlm = maxfdij g where the max is taken
over every (i; j ) for which gij  0. If no such pair exists, then go to 8.
28 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
(v) Randomly select a grid point k on one of the shortest ways from l to
m, such that, j dlk ? dmk j  1. Then, choose randomly,
2 [0; z < w]
where w and z are the input parameters to the algorithm.
(vi) Set flm :=
; flk := flk + (w ?
); fmk := fmk + (w ?
) and glm :=
glk := gmk := 1.
(vii) Endwhile.
(viii) Finally, generate the random permutation p = p (i); i = 1; 2; : : : ; n,
where P  will be the optimal permutation. Form the matrix F = (cij )
in which fij = fuv where i = p (u) and j P
= pP (v).
(ix) Output F , D and p and optimal cost w( dij ).

Next, we provide the proof of correctness of the above algorithm [165]. Be-
fore applying Step #8, the identity permutation is an optimal permutation
for the P
QAP
P
with input data, the matrices F and D and the optimal cost,
C = w( dij ).

The proof is by induction on the number of iterations.

Base: To start with, let F (0) = (fij = w) be the ow matrix. Obviously, for
this F (0) all permutations are optimal and the corresponding cost is C .
Hypothesis: At the end of the i-th iteration, for the ow matrix F (i), the
identity permutation is optimal and the corresponding cost is C .
To prove: at the end of the (i + 1)-th iteration, for the ow matrix F (i+1),
the identity permutation is optimal and the corresponding cost is the same as
at the end of the i-th iteration i.e., C .
During the (i + 1)-th iteration, let dlm = maxfdij g over all i; j such that
gij  0. Also, let a grid point k be selected such that k is on the shortest
path between l and m, i.e. dlk + dmk = dlm (Comment: k need not satisfy
the condition j dlk ? dmk j  1 as stated in the algorithm. It is sucient if
dlk + dmk = dlm ). Then,
F (i+1) = F (i) + F^
where F^ = (f^ij ) and
8
>
>
w ?
if i = l and j = k;
f^ij = > w
??
if i = m and j = k;
<

>
:
w if i = l and j = m;
0 otherwise.
It is obvious that
 THE QUADRATIC ASSIGNMENT PROBLEM 29

opt(F (i+1) )  opt(F (i) ) + opt(F^ ):

From the hypothesis opt(F (i)) = C and it can be easily proved that opt(F^ ) =
0. Hence,
opt(F (i+1) )  C:
But, again from the hypothesis, the identity permutation is optimal for F (i)
and it can be checked easily to show that it is also optimal for F^ . Hence, the
identity permutation must be optimal for F (i+1) and the optimal cost must be
the same as at the end of the i-th iteration, i.e. C .

Next, we provide two test problems constructed using the above generator.

Example 1: Here, n = 10 and the grid dimensions are: r = 5 and s = 2.

Also, the input parameters to the generator are: w = 9 and z = 3. The generated
matrices F and D are as follows:
2
9 3 2 1 16 3 41 9 0 32 3
6 15 9 0 16 18 2 3 9 40 31 77
6
6 0 2 9 22 58 9 43 0 1 3 77
6
6 1 1 0 9 15 9 33 1 17 25 77
6

F = 666 170 00 19 352 93 29 30 1 16 18 777

6

6
39 3 38 3 77
6 27 2 9 15 16 3 9 2 16 3 77
6
6 33 9 0 3 40 0 18 9 15 17 77
6
4 16 2 3 15 15 3 17 2 9 25 5
30 9 0 29 18 1 3 0 9 9
2
0 1 1 2 2 3 3 4 4 53
6 1 0 2 1 3 2 4 3 5 4 7
6 7
6 1 2 0 1 1 2 2 3 3 4 7
6 7
6 2 1 1 0 2 1 3 2 4 3 7
6 7

D = 666 32 23 21 12 10 01 21 12 32 23 777
6 7

6 7
6 3 4 2 3 1 2 0 1 1 2 7
6 7
6 4 3 3 2 2 1 1 0 2 1 7
6 7
4 4 5 3 4 2 3 1 2 0 1 5
5 4 4 3 3 2 2 1 1 0
For this problem, the known optimal cost is 1890 and the optimal permutation
of the facilities in relation to the locations is (8 2 1 10 5 9 7 4 3 6). The results
obtained by our algorithm are: cost due to our algorithm is 1890, which is
30 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
same as the optimal value and the corresponding permutation of the facilities is
(7 10 1 4 5 2 3 9 6 8).

Example 2: The grid for this example is r = 2, s = 5 and n = 10. The input
parameters to the generator are: w = 9 and z = 5. The generated ow matrix
F is given below.
2
9 21 29 17 1 17 20 16 2 1 3
6 9 9 18 34 2 2 34 14 0 4 77
6
6 5 13 9 15 3 2 16 24 18 2 77
6
6 5 26 20 9 9 4 2 2 27 3 77
6

F = 666 91 190 475 224 09 90 275 262 27 31 26 777

6

6
1 77
6 1 24 22 2 3 9 9 9 24 5 7
6 7
6 1 0 16 4 1 17 18 9 21 17 7
6 7
4 5 0 5 18 0 5 33 13 9 18 5
2 0 0 26 9 4 2 41 35 9
2
0 1 2 3 4 1 2 3 4 53
6 1 0 1 2 3 2 1 2 3 4 7
6 7
6 2 1 0 1 2 3 2 1 2 3 7
6 7
6 3 2 1 0 1 4 3 2 1 2 7
6 7

D = 666 14 23 32 41 50 05 14 23 32 41 777
6 7

6 7
6 2 1 2 3 4 1 0 1 2 3 7
6 7
6 3 2 1 2 3 2 1 0 1 2 7
6 7
4 4 3 2 1 2 3 2 1 0 1 5
5 4 3 2 1 4 3 2 1 0
Optimal cost for this example is also 1890. Our algorithm cost is the same as the
optimal cost and the corresponding permutation in both the cases is the identity
permutation.
6.2. Li & Pardalos Generator for QAPs with a Known Solution. Li
and Pardalos [150] have generalized the results of Palubetskis and constructed
test problems for more general types of QAPs. Their generator includes Palubet-
skis' procedure [175] as a special case, in which the distance matrix is taken from
a grid graph. The fortran generator described in [150] is available by e-mail from
the authors (the fortran code can also be obtained by sending an e-mail message
to \coap@math.u .edu", and in the body of the message put \send 92006").
6.3. QAP-LIB. Finally we point out that a collection of more than 130 in-
stances of quadratic assignment problems is contained in a library called the
"QAP-LIB", [39]. This library consists of two parts. The data part contains
various instances given by the input matrices A, B and C , if C 6= 0. Then
 THE QUADRATIC ASSIGNMENT PROBLEM 31
there is a documentation, corresponding to [39]. This documentation is updated
regularly, the last update being from February 1994. The documentation con-
tains the following information on each of the instances: the best known feasible
solution value is given, along with some information on who found it and by
which method. Secondly the best currently available lower bounds on the objec-
tive function is provided. The two parts are available via anonymous ftp from
ftp.tu-graz.ac.at in the directory /pub/papers/qaplib.

6.4. OR-Library. QAP instances can be also obtained from the OR-Library
(o.rlibrary@ic.ac.uk) - see the le qapinfo. For details see [123]. Information
about test problems for quadratic assignment problems as well as other combi-
natorial problems can be obtained by sending email to o.rlibrary@ic.ac.uk with
the email message being the le name for the problem areas you are interested
in.

7. Concluding Remarks
In this paper we gave a survey regarding the most recent results and applica-
tions on QAP. In addition, an up-to-date bibliography is included which includes
papers on QAP and realted problems as well as applications in diverse areas.
32 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
Appendix A. Notations
n The size of an instance of the QAP
e The column vector of ones
A The ow matrix
B The distance matrix
QAP (A; B ) An instance of the QAP with ow matrix A and distance matrix B
m The set of permutations
f (p) The objective function value of QAP (A; B ) corresponding to p
0
fA;B The maximum objective function value of QAP (A; B )

fA;B The minimum objective function value of QAP (A; B )
GLB (A; B ) A permutation yielding the maximum objective function value of QAP (A; B )
EV B (A; B ) The Eigenvalue bound for the QAP (A; B )
hx; yi The inner product of vectors x and y
hx; yi+ The maximum permuted inner product of vectors x and y
hx; yi? The minimum permuted inner product of vectors x and y
x+ The vector obtained by reordering the components of x ascendingly
x? The vector obtained by reordering the components of x descendingly
At The transpose of the matrix A
trace A The trace of the matrix A
diag (A) The vector formed from the diagonal elements of A
diag (v) The diagonal matrix formed from the vector v
 The set of permutation matrices
P The set of positive semide nite symmetric matrices
D The set of doubly stochastic matrices
O The set of orthogonal matrices
N The set of nonnegative (elementwise) matrices
E The set of matrices with row and column sums 1
S The set of matrices with the sum of the squares of all elements equal to n

References
1. E. Aarts and J. Korst, Simulated annealing and Boltzman machines: A stochastic ap-
proach to combinatorial optimization and neural computing, John Wiley and Sons, 1989.
2. W.P. Adams and H.D. Sherali, A tight linearization and an algorithm for zero-one
quadratic programming problems, Management Science 32 (1986), no. 10, 1274{1290.
3. R.K. Ahuja, T.L. Magnanti, and J.B. Orlin, Some recent advances in network ows,
SIAM Review 33 (1991), no. 2, 175{219.
4. G.C. Armour and E.S. Bu a, Heuristic algorithm and simulation approach to relative
location of facilities, Management Science 9 (1963), 294{309.
5. A.A. Assad and W. Xu, On lower bounds for a class of quadratic f0 1g programs, Oper-
;

ations Research Letters 4 (1985), 175{180.

6. J.G. Augustin and J. Minker, Analysis of some graph theoretical cluster techniques, J.
ACM 17 (1970), 571{588.
7. G. Avondo-Bodeno, Economic applications of the theory of graphs, Gordon and Breach
Science Publishers, 1962.
8. E. Balas and J.B. Mazzola, Quadratic 0-1 programming by a new linearization, Proceed-
ings of the TIMS/ORSA (Washington D.c.), TIMS/ORSA, May 1980.
 THE QUADRATIC ASSIGNMENT PROBLEM 33
9. E. Balas and L. Qi, Linear-time separation algorithms for the three-index assignment
polytope, Discrete Applied Mathematics (1993), 1{12.
10. E. Balas and M.J. Saltzman, Facets of the three-index assignment polytope, Discrete
Applied Mathematics (1989), 201{229.
11. , An algorithm for the three-index assignment problem, Operations Research 39
(1991), 150{161.
12. E. Balas and C.S. Yu, Finding a maximum clique in an arbitrary graph, SIAM Journal
of Computing 15 (1986), 1054{1068.
13. H.-J. Bandelt, Y. Crama, and F.C.R. Spieksma, Approximation algorithms for multidi-
mensional assignment problems with decomposable costs, Discrete Applied Mathematics
(1991), to appear.
14. J.F. Bard and T.A. Feo, Operations sequencing in discrete parts manufacturing, Man-
agement Science 35 (1989), 249{255.
15. J.F. Bard and T.A. Feo, An algorithm for the manufacturing equipment selection problem,
IIE Transactions 23 (1991), 83{92.
16. M.S. Bazaraa and A.N. Elshafei, An exact branch-and-bound procedure for the quadratic-
assignment problem, Naval Research Logistics Quarterly 26 (1979), 109{121.
17. M.S. Bazaraa and O. Kirca, A branch-and-bound heuristic for solving the quadratic as-
signment problem, Naval Research Logistics Quarterly 30 (1983), 287{304.
18. M.S. Bazaraa and H.D. Sherali, New approaches for solving the quadratic assignment
problem, Operations Research Verfahren 32 (1979), 29{46.
19. , Bender's partitioning scheme applied to a new formulation of the quadratic
assignment problem, Naval Research Logistics Quaterly 27 (1980), 29{41.
20. , On the use of exact and heuristic cutting plane methods for the quadratic as-
signment problem, Journal of Operations Research Society 33 (1982), 991{1003.
21. G. Birko , Tres observaciones sobre el algebra lineal, Univ. Nac. Tucuman Rev. Ser. A
(1946), 147{151.
22. J. B. Blanks, Near optimal quadratic-based placement for a class of ic layout problems,
IEEE Circuits and Devices (1985), 31{373.
23. S.H. Bokhari, On the mapping problem, IEEE Transactions on Computers C-30 (1981),
no. 3, 207{214.
24. , Assignment problems in parallel and distributed computing, Kluwer Academic
Publishers, Boston, 1987.
25. V. Joseph Bowman and Fred Glover, A note on zero-one integer and concave program-
ming, Operations Research 21 (1973), 182{183.
26. S.C. Boyd, The subtour polytope of the travelling salesman problem, Ph.D. thesis, Uni-
versity of Waterloo, 1986.
27. E. Donald Brown, L. Christopher Huntley, and R. Andrew Spillance, A parallel genetic
heuristic for the quadratic assignment problem, Proc. 3rd Conf. on Genetic Algorithms,
406-415, 1989.
28. R.A. Brualdi and H.J. Ryser, Combinatorial matrix theory, Cambridge University Press,
New York, 1991.
29. P.A. Bruijs, On the quality of heuristic solutions to a 19 x 19 quadratic assignment
problem, European Journal of Operational Research 17 (1984), 21{30.
30. E.S. Bu a, G.C. Armour, and T.E. Vollmann, Allocating facilities with CRAFT, Harvard
Business Review 42 (1962), 136{158.
31. R.E. Burkard, Die st rungsmethode zur l sung quadratischer zuordnungsprobleme, Oper-
o o

ations Research Verfahren 16 (1973), 84{108.

32. , Some recent advances in quadratic assignment problems, MathematicalProgram-
ming, Elsevier Publishers B.V. North-Holland, 1984, pp. 53{68.
33. R.E. Burkard, Locations with spatial interactions: the quadratic assignment problem,
Discrete Location Theory (P.B. Mirchandani and R.L. Francis, eds.), John Wiley, 1991.
34. R.E. Burkard and T. Bonniger, A heuristic for quadratic boolean programs with applica-
tions to quadratic assignment problems, European Journal of Operational Research 13
(1983), 374{386.
34 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
35. R.E. Burkard, E. Cela, and B. Klinz, On the biquadratic assignment problem, Tech.
Report 260, Institute of Mathematics, University of Technology Graz, 1993.
36. R.E. Burkard and U. Derigs, Assignment and matching problems: Solution methods
with fortran programs, Lecture Notes in Economics and Mathematical Systems, vol. 184,
Springer, Berlin, 1980.
37. R.E. Burkard and U. Fincke, The asymptotic probabilistic behaviour of quadratic sum
assignment problems, Z. Operations Res. 27 (1983), 73{81.
38. R.E. Burkard and K. Frohlich, Some remarks on three-dimensional assignment problem,
Method of Oper. Res. 36 (1980), 31{36.
39. R.E. Burkard, S. Karisch, and F. Rendl, QAPLIB { a quadratic assignment problem
library, European Journal of Operations Research 55 (1991), 115{119, Updated version
{ Feb. 1993.
40. R.E. Burkard and J. O ermann, Entwurf von schreibmaschinentastaturen mittels
quadratischer zuordnungsprobleme, Z. Operations Res. 21 (1977), B121{B132.
41. R.E. Burkard and F. Rendl, A thermodynamically motivated simulation procedure for
combinatorial optimization problems, European Journal of Operations Research 17
(1984), 169{174.
42. R.E. Burkard and R. Rudolf, Computational investigations on 3-dimensional axial as-
signment problems, Tech. report, Technische Universitat Graz, Austria, 1993, to appear
in: Belgian Journal of Operations Research.
43. R.E. Burkard, R. Rudolf, and G. Woeginer, Three-dimensional axial assignment prob-
lems with decomposable cost-coecients, Tech. Report Report 238, Technische Universitat
Graz, Austria, 1992.
44. R.E. Burkard and W. Sandholzer, Eciently solvable special cases of bottleneck travelling
salesman problems, Discrete Applied Mathematics 32 (1991), 61{76.
45. R.E. Burkard and K.-H. Stratmann, Numerical investigations on quadratic assignment
problems, Naval Research Logistics Quaterly 25 (1978), 129{148.
46. P. Camion, Characterization of totally unimodular matrices, Proc. Amer. Math. Soc. 16
(1965), 1068{73.
47. P. Carraresi and F. Malucelli, A new lower bound for the quadratic assignment problem,
Operations Research 40 (1992), no. Supplement 1, S22{S27.
48. P. Carraresi and F. Malucelli, A reformulation and new lower bounds for the quadratic
assignment problem, Tech. Report TR-34/92, Universita di Pisa, 1992.
49. J. Chakrapani and J. Skorin-Kapov, A connectionist approaches to the quadratic assign-
ment problem, Computers and Operations Research 19 (1992), no. 3/4, 287{295.
50. , A constructive method to improve lower bounds for the quadratic assignment
problem, Working paper, State University of New York at Stony Brook, 1992.
51. , Massively parallel tabu search for the quadratic assignment problem, Annals of
Operations Research forthcoming (1992).
52. N. Christo des and E. Benavent, An exact algorithm for the quadrtic assignment problem,
Operations Research 37 (1989), no. 5, 760{768.
53. N. Christo des and M. Gerrard, Special cases of the quadratic assignment problem, Man-
agement Science Research Report 391, Carnegie Mellon University, April 1976.
54. , A graph theoretic analysis of bounds for the quadratic assignment problem, Stud-
ies on graphs and discrete programming (P. Hansen, ed.), North-Holland, 1981, pp. 61{68.
55. N. Christo des, A. Mingozzi, and P. Toth, Contributions to the quadratic assignment
problem, European Journal of Operations Research 4 (1980), 243{247.
56. A. Rogger C.N. Fiechter and D. de Werra, Basic ideas of tabu search with an application
to traveling salesman and quadratic assignment, Ricerca Operativa 62 (1992), 5{28.
57. D.T. Connolly, An improved annealing scheme fro the qap, Journal of Operational Re-
search 46 (1990), 93{100.
58. K. Conrad, Das quadratische zuweisungsproblem und zwei seiner spezialfalle, Mohr-
Siebeck, Tubingen (1971).
59. Y. Crama and F.C.R. Spieksma, Approximation algorithms for three-dimensional assign-
ment problems with triangle inequalities, European Journal of Operational Research 60
 THE QUADRATIC ASSIGNMENT PROBLEM 35
(1992), 273{279.
60. P.J. Davis, Circulant matrices, John Wiley & Sons, New York, 1979.
61. N. Deo, Graph theory with applications to engineering and computer science, Prentice-
Hall, 1974.
62. J.W. Dicky and J.W. Hopkins, Campus building arrangement using TOPAZ, Transporta-
tion Research 6 (1972), 59{68.
63. R.E. Echols and L. Cooper, Solution of integer linear programming problems by direct
search, J. Assoc. Comput. Mach. 15 (1968), 75{84.
64. P.M. Pardalos (Editor), Complexity in numerical optimization, World Science, 1993.
65. C.S. Edwards, The derivation of a greedy approximator for the koopmans-beckmann
quadratic assignment problem, Proc. CP77 Combinatorial Prog. Conf. (1977), 55{86.
66. , A branch and bound algorithm for the Koopmans-Beckman quadratic assignment
problem, Mathematical Programming Study 13 (1980), 35{52.
67. A.N. Elshafei, Hospital layout as a quadratic assignment problem, Operations Research
Quarterly 28 (1977), 167{179.
68. R. Euler, Odd cycles and a class of facets of the axial 3-index assignment polytope,
Applicationes Mathematicae (Zastosowania Matematyki) XIX (1987), 375{386.
69. B.H. Faaland and F.S. Hillier, Interior path methods for heuristic integer programming
procedures, Operations Research 27 (1979), 1069{1087.
70. K. Fan, Maximum properties and inequalities for the eigenvalues of completely continuous
operators, Proc. Nat. Acad. Sci. U.S.A. 35 (1951), 1951.
71. T.A. Feo and J. Gonzalez-Velarde, The intermodal trailer assignment problem, Tech.
report, Operations Research Group, The University of Texas at Austin, Austin, TX
78712-1063, 1992.
72. T.A. Feo and M.G.C. Resende, A probabilistic heuristic for a computationally dicult
set covering problem, Operations Research Letters 8 (1989), 67{71.
73. T.A. Feo and M.G.C. Resende, Greedy randomized adaptive search procedures, Tech.
report, AT&T Bell Laboratories, Murray Hill, NJ 07974-2070, 1994.
74. G. Finke, R.E. Burkard, and F. Rendl, Quadratic assignment problems, Annals of Discrete
Mathematics 31 (1987), 61{82.
75. G. Finke and E.B. Medova-Dempster, Approximation approach to combinatorial opti-
mization problems, Tech. report, Technical University of Nova Scotia, 1987.
76. C.A. Floudas and P.M. Pardalos, A collection of test problems for constrained global
optimization algorithms, Springer-Verlag, Lecture Notes in Computer Science, No. 455,
1990.
77. R.L. Francis and J.A. White, Facility layout and location, Prentice-Hall, Englewood Cli s,
N.J., 1974.
78. J.C.B. Frenk, M. van Houweninge, and A.H.G. Rinnoony Kan, Asymptotic properties of
assignment problems, Tech. report, Erasmus University, Rotterdam, 1982.
79. C. Friden, A. Hertz, and D. de Werra, STABULUS: A technique for nding stable sets
in large graphs with tabu search, Computing 42 (1989), 35{44.
80. A.M. Frieze, Complexity of a 3-dimensional assignment problem, European Journal of
Operations Research 13 (1983), 161{164.
81. A.M. Frieze and J. Yadegar, an algorithm for solving 3-dimensional assignment problem
with application to scheduling a teaching practice, Operations Research 32 (1981), 989{
995.
82. , On the quadratic assignment problem, Discrete Applied Mathematics 5 (1983),
89{98.
83. K. Frohlich, Dreidimensionale zuordnungsprobleme, Ph.D. thesis, Math. Institut, Univer-
sitat Koln, 1979, Masters Thesis.
84. W. Gander, G. Golub, and U. von Matt, A constrained eigenvalue problem, Linear Alge-
bra and its Applications 114/115 (1989), 815{839.
85. M.R. Garey and D.S. Johnson, Computers and intractability - A guide to the theory of
NP-completeness, W.H. Freeman and Company, 1979.
86. J.W. Gavett and N.V. Plyter, The optimal assignment of facilities to locations by branch
36 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
and bound, Operations Research 14 (1966), 210{232.
87. A.M. Geo rion, An improved implicit enumeration approach for integer programming,
Operations Research 17 (1969), 437{454.
88. A.M. Geo rion and G.W. Graves, Scheduling parallel production lines with changeover
costs: Practical applications of a quadratic assignment/LP approach, Operations Re-
search 24 (1976), 595{610.
89. P.C. Gilmore, Optimal and suboptimal algorithms for the quadratic assignment problem,
J. SIAM 10 (1962), 305{313.
90. F. Glover, Tabu search - Part 1, ORSA Journal on Computing 1 (1989), no. 3, 190{206.
91. , Tabu search - Part 2, ORSA Journal on Computing 2 (1989), no. 1, 4{32.
92. Fred Glover and D. Klingman, Concave programming applied to a special class of 0-1
integer programs, Operations Research 21 (1973), 135{140.
93. E. David Goldburg, Genetic algorithms in search, optimization, and machine learning,
Addison-Wesley Publishing Company, Inc., 1989.
94. B. Gollan, Eigenvalue perturbations and nonlinear parametric optimization, Mathemati-
cal Programming Study 30 (1987), 67{81.
95. S. Goto and E.S. Kuh, An approach to the two-dimensional placement problem in circuit
layout, IEEE Transactions on Circuits and Systems 25 (1978), no. 4, 208{214.
96. A. Graham, Kronecker products and matrix calculus: with applications, Halsted Press,
Toronto, 1981.
97. G.W. Graves and A.B. Whinston, An algorithm for the quadratic assignment problem,
Management Science 17 (1970), 453{471.
98. V.P. Gulati, S.K. Gupta, and A.K. Mittal, Unconstrained bivalent programming problem,
European Journal of Operations Research 15 (1984), 121{125.
99. G. Gwan and L. Qi, On facet of the three index assignment polytope, Australasian Journal
of Combinatorics 6 (1992), 67{87.
100. S.W. Hadley, Continuous optimization approaches for the quadratic assignment problem,
Ph.D. thesis, University of Waterloo, 1989.
101. S.W. Hadley, F. Rendl, and H. Wolkowicz, Bounds for the quadratic assignment prob-
lem using continuous optimization techniques, Integer Programming and Combinatorial
Optimization, University of Waterloo Press, 1990, pp. 237{248.
102. , A new lower bound via projection for the quadratic assignment problem, Math-
ematics of Operations Research 17 (1992), no. 3, 727{739.
103. S.W. Hadley, F. Rendl, and H. Wolkowicz, Nonsymmetric quadratic assignment problems
and the Ho man-Wielandt inequality, Linear Algebra and its Applications 58 (1992),
109{124.
104. M. Hall Jr., Combinatorial theory, Blaisdell Company, Waltham, MA, 1967.
105. M. Hanan and J.M. Kurtzberg, A review of the placement and quadratic assignment
problem, SIAM Review 14 (1972), 324{342.
106. P. Hansen and L. Kaufman, A primal-dual algorithm for the three dimensional assignment
problem, Cahiers Centre E tudess Rech. Oper 15 (1973), 327{336.
107. J.P. Hart and A.W. Shogan, Semi-greedy heuristics: An empirical study, Operations
Research Letters 6 (1987), 107{114.
108. D.R. Heey, Assigning runners to a relay team, Optimal Strategies in Sports (Amster-
dam) (S.P. Ladany and R.E. Machol, eds.), North-Holland, Amsterdam, 1977, pp. 169{
171.
109. C.H. Heider, A computationally simpli ed pair exchange algorithm for the quadratic as-
signment problem, Paper 101, Center for Naval Analysis, Arlington (Va), 1972.
110. C. Helmberg, B. Mohar, S. Poljak, and F. Rendl, A spectral approach to bandwidth
and separator problems in graphs, Tech. Report CDLDO 32, Institute of Mathematics,
University of Technology Graz, 1993.
111. P.S. Hiller and M.M. Connors, Quadratic assignment algorithms and the location of in-
divisible facilities, Management Science 13 (1966), 42{57.
112. F.S. Hillier, Ecient heuristic procedures for integer linear programming with an interior,
Operations Research 17 (1969), 600{637.
 THE QUADRATIC ASSIGNMENT PROBLEM 37
113. , A further investigation of ecient heuristic procedures for integer linear pro-
gramming with an interior, Tech. report, Dept. of Operations Research, Stanford Uni-
versity, Stanford, CA 94305, 1977.
114. J.B. Hiriart-Urruty, From convex optimization to nonconvex optimization. part i: Nec-
essary and sucient conditions from global optimality, Nonsmooth Optimization and
Related Topics (F.H. Clarke, ed.), Plenum Press, 1989, pp. 219{239.
115. A.J. Ho man and H.W. Wielandt, The variation of the spectrum of a normal matrix,
Duke Mathematics 20 (1953), 37{39.
116. K.H. Ho mann and P. Salamon, Simulated annealing for single minimum optimization
problems, International J. Computer Math. 39 (1991), 193{204.
117. R. Horn and C. Johnson, Matrix analysis, Cambridge University Press, New York, 1985.
118. L.J. Hubert, Assignment methods in combinatorial data analysis, Marcel Dekker, Inc.,
New York, NY 10016, 1987.
119. T. Ibaraki, Theoretical comparisons of search strategies in branch-and-bound algorithms,
International Journal of Computer and Information Sciences 5 (1976), no. 4, 315{344.
120. T. Ibaraki, T. Ohashi, and F. Mine, A heuristic algorithm for mixed-integer programming
problems, Mathematical Programming Study 2 (1974), 115{136.
121. I.Gilar and M.A. Pollatschek, Layout simulation for keyboard, Behaviour and Information
Technology 5 (1986), 273{281.
122. B. Jansen, A note on \Lower bounds for the QAP", Tech. report, Delft University of
Technology, Mathematics and Computer Science, December 1993.
123. J.E.Beasley, Obtaining test problems using e-mail, Journal of global optimization (1994),
to appear.
124. Prashnna Jog, Jung Y. Suh, and Dirk Van Gucht, Parallel genetic algorithms applied to
the traveling salesman problem, SIAM J. Optimization 1 (1991), no. 4, 515{529.
125. D.S. Johnson, C.A. Aragon, L.A. McGeoch, and C. Schevon, Optimization by simulated
annealing: An experimental evaluation; Part II, Graph coloring and number partitioning,
Operations Research 39 (1991), 378{406.
126. D.S. Johnson, C.H. Papadimitriou, and M. Yannakakis, How easy is local search?, Journal
of Computer and System Sciences 37 (1988), 79{100.
127. B. K. Kaku and G. L. Thompson, An exact algorithm for the general quadratic assignment
problem, European Journal of Operational Research 23 (1986), 382{390.
128. B.K. Kaku, T.E. Morton, and G.L. Thompson, A heuristic algorithm for the facilities
layout problem, Tech. report, Carnegie Mellon University, Pittsburgh, Pa., 1989.
129. S. E. Karisch, F. Rendl, and H. Wolkowicz, Trust regions and the quadratic assignment
problem, Research Report CORR 93-15, DIMACS, Rutgers University, New Brunswick,
NJ, 1993, To appear in the Proceedings of the Workshop on the Quadratic Programming
Problem, DIMACS, 1993.
130. N. Karmarkar, An interior-point approach to NP-complete problems { extended abstract,
Contemporary Mathematics 114 (1990), 297{308.
131. R. Karp, Reducibility among combinatorial problems, Proc. Complexity of Computer
Computations (R.E. Miller and J.W. Thatcher, eds.), Plenum Press, 1972, pp. 85{104.
132. L. Kaufman and F. Broeckx, An algorithm for the quadratic assignment problem using
benders' decomposition, European Journal of Operational Research 2 (1978), 204{211.
133. B. Kernighan and S. Lin, An ecient heuristic procedure for partitioning graphs, Bell
Systems Journal 49 (1972), 291{307.
134. G.A.P. Kindervater and J.K. Lenstra, An introduction to parallelism in combinatorial
optimization, Discrete Applied Mathematics 14 (1986), 135{156.
135. S. Kirpatrick, C.D. Gelatti, and M.P. Vecchi, Optimization by simulated annealing, Sci-
ence 220 (1983), 671{680.
136. G.A. Kochenberger, B.A. McCarl, and F.P. Wyman, A heuristic for general integer pro-
gramming, Decision Sci. 5 (1974), 36{44.
137. T.C. Koopmans and M.J. Beckmann, Assignment problems and the location of economic
activities, Econometrica 25 (1957), 53{76.
138. J. Krarup and P.M. Pruzan, Computer-aided layout design, Mathematical Programming
38 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
Study 9 (1978), 75{94.
139. T.A. Lacksonen and Jr E.E. Enscore, Quadratic assignment algorithms for the dynamic
layout, Int. J. Prod. Res. 31 (1993), 503{517.
140. M. Laguna and J.L. Gonzalez-Velarde, A search heuristic for just-in-time scheduling in
parallel machines, Journal of Intelligent Manufacturing 2 (1991), 253{260.
141. A. M. Land, A problem of assignment with interrelated costs, Operations Research Quar-
terly 14 (1963), 185{198.
142. R. S. Lashkari and S. C. Jaisingh, An heuristic approach to quadratic assignment problem,
Journal of the Operational Research Society 31 (1980), no. 9, 845{850.
143. E.L. Lawler, The quadratic assignment problem, Management Science 9 (1963), 586{599.
144. E.L. Lawler, J.K. Lenstra, A.H.G. Rinnooy Kan, and D.B. Shmoys, The traveling sales-
man problem: A guided tour of combinatorial optimization, John Wiley & Sons, 1985.
145. O. Leue, Methoden zur losung dreidimensionaler zuordnungsprobleme, Angew. Inform.
(1972), 154{162.
146. Tao Li, Parallel imprecise iterative deepening for combinatorial optimization, Interna-
tional J. of High Speed Computing 3 (1991), no. 1, 63{76.
147. Y. Li, Heuristic and exact algorithms for the quadratic assignment problem, Ph.D. thesis,
The Pennsylvania State University, 1992.
148. Y. Li, P.M. Pardalos, K.G. Ramakrishnan, and M.G.C. Resende, A branch-and-bound
algorithm for the quadratic assignment problem, Tech. report, AT&T Bell Laboratories,
Murray Hill, NJ 07974-2070, December 1992.
149. , Lower bounds for the quadratic assignment problem, Tech. report, AT&T Bell
Laboratories, Murray Hill, NJ 07974-2070, April 1992, To appear in Annals of Operations
Research.
150. Yong Li and Panos M. Pardalos, Generating quadratic assignment test problems with
known optimal permutations, Computational Optimization and Applications 1 (1992),
no. 2, 163{184.
151. , Parallel algorithms for the quadratic assignment problem, in Recent Advances in
Optimization and Parallel Computing, 177{189, Elsevier, Amsterdam, 1992, pp. 177{189.
152. S. Lin and B. Kernighan, An e ective heuristic algorithm for the travelling salesman
problem, Operations Research 21 (1973), 498{516.
153. L. Lovasz, On the ratio of optimal integral and fractional covers, Discrete Mathematics
13 (1975), 391{398.
154. R. F. Love and J. Y. Wong, On solving a one-dimensional space allocation problem with
integer programming, INFOR 14 (1976), 139{143.
155. , Solving quadratic assignment problems with rectangular distances and integer
programming, Naval Research Logistics Quarterly 23 (1976), 623{627.
156. Robert F. Love, James G. Morris, and George O. Wesolowsky, Facilities location: Models
and methods, North-Holland, 1988.
157. M. Grotschel and L. Lovasz and A. Schrijver, Geometric algorithms and combinatorial
optimization, Springer-Verlag, 1988.
158. W.L. Maxwell, The scheduling of economic lot sizes, Naval Research Logistics Quaterly
11 (1964), 89{124.
159. E.J. McCormik, Human factors engineering, McGraw-Hill, New York, 1970.
160. P.B. Mirchandani and T. Obata, Locational decisions with interactions between facil-
ities: the quadratic assignment problem a review, Working Paper Ps-79-1, Rensselaer
Polytechnic Institute, Troy, New York, May 1979.
161. H. Muhlenbein, Parallel genetic algorithms, population genetics and combinatorial opti-
mization, Proc. 3rd Conf. on Genetic Algorithms, 416-421, 1989.
162. , Parallel genetic algorithms in optimization, to appear in informatik fachbenchle,
Springer-Verlag, 1992.
163. H. Muhlenbein, M. Gorges-Schleuter, and O. Kramer, Evolution algorithms in combina-
torial optimization, Parallel Computing 7 (1988), 65{85.
164. H. Muller-Merbach, Optimale reihenlorgen, Springer, Berlin, 1970.
165. K.A. Murthy and P.M. Pardalos, A polynomial-time approximation algorithm for the
 THE QUADRATIC ASSIGNMENT PROBLEM 39
quadratic assignment problem, Tech. Report CS-33-90, The Pennsylvania State Univer-
sity, July 1990.
166. K.A. Murthy, P.M. Pardalos, and Y. Li, A local search algorithm for the quadratic as-
signment problem, Informatica 3 (1992), no. 4, 524{538.
167. L.D. Muu and W. Oettli, Method for minimizing a convex-concave function over a convex
set, Journal of Optimization and Applications 70 (1991), no. 2, 377{384.
168. M.A. Pollatschek N. Gershoni and Y.T. Radday, A better hebrew typewritter?, Industrial
Engineering October (1976), 35{37.
169. George L. Nemhauser and Laurence A. Wolsey, Integer and combinatorial optimization,
John Wiley and Sons, 1988.
170. V. Nissen, A new ecient evolutionary algorithm for the quadratic assignment prob-
lem, Operations Research Proceedings (K.W. Hansmann, A. Bachem, M. Jarke, W.E.
Katzenberg, and A. Marusev, eds.), Springer Berlin, 1992, pp. 259{267.
171. C.E. Nugent, T.E. Vollmann, and J. Ruml, An experimental comparison of techniques
for the assignment of facilities to locations, Journal of Operations Research 16 (1969),
150{173.
172. M.L. Overton, Large-scale optimization of eigenvalues, SIAM J. Optimization 2 (1992),
88{120.
173. M. Padberg, On the facial structure of the set packing polyhedra, Mathematical Program-
ming 5 (1973), 199{216.
174. G. Palubetskis, Quadratic 0-1 optimization, Informatica 1 (1990), 89{106.
175. G.S. Palubetskis, Generation of quadratic assignment test problems with known optimal
solutions (in Russian), Zh. Vychisl. Mat. Mat. Fiz. 28 (1988), no. 11, 1740{1743.
176. C.H Papadimitriou and K. Steiglitz, Combinatorial optimization: algorithms and com-
plexity, Prentice-Hall, Inc., Englewood Cli s, NJ 07632, USA, 1982.
177. C.H. Papadimitriou and D. Wolfe, The complexity of facets resolved, Proceedings of the
Foundations Of Computer Science, 1985, pp. 74{78.
178. P.M. Pardalos, Generation of large-scale quadratic programs for use as global optimization
test problems, ACM Transactions on Mathematical Software 13 (1987), no. 2, 133{137.
179. , Construction of test problems in quadratic bivalent programming, ACM Trans-
actions on Mathematical Software 17 (1991), no. 1, 74{87.
180. P.M. Pardalos and J. Crouse, A parallel algorithm for the quadratic assignment problem,
Proceedings of the Supercomputing 1989 Conference, ACM Press, 1989, pp. 351{360.
181. P.M. Pardalos and S. Jha, Complexity of uniqueness and local search in quadratic 0-1
programming, Operations Research Letters (1992), 119{123.
182. P.M. Pardalos and X. Li, Parallel branch and bound algorithms for combinatorial opti-
mization, Supercomputer 39 (1990), 23{30.
183. P.M. Pardalos, K.A. Murthy, and T.P. Harrison, A computational comparison of local
search heuristics for solving quadratic assignment problems, Informatica 4 (1993), no. 1-
2, 172{187.
184. P.M. Pardalos, K.A. Murthy, and Y. Li, Computational experience with parallel algo-
rithms for solving the quadratic assignment problem, to appear in Computer Science and
Operations Research: New developments in their interfaces (Williamsburg, VA), ORSA
CSTS, January 1992.
185. P.M. Pardalos, A. Phillips, and J.B. Rosen, Topics in parallel computing in mathematical
programming, Science Press, 1993.
186. P.M. Pardalos and G.P. Rodgers, Parallel branch and bound algorithms for uncon-
strained quadratic 0-1 programming, Impact of Recent Advances on Operations Research
((R. Sharda et al.), ed.), North-Holland Press, 1989, pp. 131{143.
187. , Computational aspects of a branch and bound algorithm for quadratic zero-one
programming, Computing 45 (1990), 131{144.
188. , A branch and bound algorithm for the maximum clique problem, Comp. and
Oper. Research 19 (1992), no. 5, 363{375.
189. P.M. Pardalos and J.B. Rosen, Methods for global concave minimization: A bibliographic
survey, Siam Review 28 (1986), 367{379.
40 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
190. , Constrained global optimization: Algorithms and applications, Lecture Notes in
Computer Science, vol. 268, Springer-Verlag, Berlin, 1987.
191. P.M. Pardalos and G. Schnitger, Checking local optimality in constrained quadratic pro-
gramming is np-hard, Operations Research Letters 7 (1988), 33{35.
192. P.M. Pardalos and S. Vavasis, Quadratic programming with one negative eigenvalue is
np-hard, Journal of Global Optimization 1 (1991), 15{23.
193. P.M. Pardalos and Jue Xue, The maximum clique problem, Tech. report, Department of
Industrial and Systems Engineering, University of Florida, Gainesville, FL, 1992.
194. C.C. Pegels, Plant layout and discrete optimizing, International Journal of Production
Research 5 (1966), 81{92.
195. J.F. Pierce and W.B. Crowston, Tree search algorithms for quadratic assignment prob-
lems, Naval Research Logistics Quaterly 18 (1971), 1{36.
196. W.P. Pierskalla, The tri-substitution method for the three-multidimensional assignment
problem, CORS J. 5 (1967), 71{81.
197. , The multidimensional assignment problem, Operations Research 16 (1968), 422{
431.
198. S. Poljak, F. Rendl, and H. Wolkowicz, General relaxations and concave extensions for
(0,1)-quadratic programming, Tech. report, University of Waterloo, 1994, in progress.
199. M.A. Pollatschek, N. Gershoni, and Y.T. Radday, Optimization of the typewritter key-
board by computer simulation, Angewandte Informatik 10 (1976), 438{439.
200. W.R. Pulleyblank, Mathematical Programming The State of the Art, 312{345, Springer-
Verlag, 1982, pp. 312{345.
201. A.K. Dewdney P.Z. Chinn, J. Chvatalova and N.E. Gibbs, The bandwidth problem for
graphs and matrices: a survey, Journal of Graph Theory 6 (1982), 223{254.
202. L. Qi, E. Balas, and G. Gwan, A new facet class and a polyhedral method for the three-
index assignment problem, Advances in Optimization (D.-Z. Du, ed.), Kluwer Academic,
1994, pp. 256{274.
203. M.J. Quinn, Designing ecient algorithms for the speedup of parallel best-bound branch-
and-bound algorithms, McGraw-Hill, New York, 1987.
204. M. Raghavachari, On connections between zero-one integer programming and concave
programming under linear constraints, Operations Research 17 (1969), 680{684.
205. G. Reinelt, The linear ordering problem: algorithms and applications, Research and Ex-
position in Mathematics, vol. 8, Heldermann Verlag Berlin, Berlin, 1985.
206. S. Reiter and D.B. Rice, Discrete optimization solution procedures for linear and nonlin-
ear integer programming problems, Management Science 12 (1966), 829{850.
207. F. Rendl, Ranking scalar products to improve bounds for the quadratic assignment prob-
lem, European Journal of Operations Research 20 (1985), 363{372.
208. F. Rendl, Quadratic assignment problems on series-parallel digraphs, Zeitschrift fur Op-
erations Research A 30 (1986), 161{173.
209. F. Rendl and H. Wolkowicz, A projection technique for partitioning the nodes of a graph,
Annals of Operations Research (1990), no. CORR 90-20, To appear in the special issue
of APMOD93 conference Budapest 1993.
210. F. Rendl and H. Wolkowicz, Applications of parametric programming and eigenvalue max-
imization to the quadratic assignment problem, Mathematical Programming 53 (1992),
63{78.
211. W.T. Rhee, A note on asymptotic properties of teh quadratic assigment problem, Oper-
ations Research Letters (1989), 197{200.
212. , Stochastic analysis of the quadratic assignment problem, Manuscript, Ohio State
University, Columbus, Ohio 43210, 1990.
213. R.H. Roth, An approach to solving linear discrete optimization problems, J. Assoc. Com-
put. Mach. 17 (1970), 300{313.
214. C. Roucairol, A ection quadratique, Ph.D. thesis, Universite Pierre et Marie Curie, Paris
VI, 1976.
215. , A reduction method for quadratic assignment problems, Operations Research
Verfahren 32 (1979), 183{187.
 THE QUADRATIC ASSIGNMENT PROBLEM 41
216. , Un nouvel algorithme pour le probleme d'a ectation quadratique, R.A.I.R.O. 13
(1979), 275{301.
217. , Du sequentiel au parallele: la recherche aborescente a la programmation quadra-
tique en variables 0 et 1, 1987, These d'Etat.
218. , A parallel branch and bound algorithm for the quadratic assignment problem,
Discrete Applied Mathematics 18 (1987), 211{225.
219. R. Rudolf, Dreidimensionale axiale zuordnungsprobleme, Ph.D. thesis, Technische Uni-
versitat Graz, Austria, 1991, Masters Thesis.
220. S. Sahni, Computationally related problems, SIAM Journal of Computing 3 (1974), 262{
279.
221. S. Sahni and T. Gonzalez, P-complete approximation problems, Journal of the Association
of Computing Machinery 23 (1976), 555{565.
222. J.E. Savage and M.G. Wloka, Parallelism in graph-partitioning, Journal of Parallel and
Distributed Computing 13 (1991), no. 3, 257{272.
223. A.A. Sch er and M. Yannakakis, Simple local search problems that are hard to solve,
a

Tech. report, AT&T Bell Laboratories, 1989.

224. L. Schrage, A more portable fortran random number generator, ACM Transactions on
Mathematical Software 5 (1979), 132{138.
225. Alexander Schrijver, Theory of linear and integer programming, John Wiley and Sons,
1986.
226. M. Scriabin and R.C. Vergin, Comparison of computer algorithms and visual based meth-
ods for plant layout, Management Science 22 (1975), 172{187.
227. R. Sedgewick, Permutation generation methods, Computing Surveys 9 (1977), no. 2,
137{164.
228. S. Senju and Y. Toyoda, An approach to linear programming with 0-1 variables, Manage-
ment Science 15 (1968), B196{B207.
229. S.I. Sergeyev, A new lower bound for the quadratic assignment problem, Zh. Vychisl.
Mat. Mat. Fiz. 27 (1987), no. 12, 1802{1811.
230. C.E. Shannon, The zero-error capacity of a noisy channel, I.R.E. Transactions 3 (1956).
231. SHARE, Computerized relative allocation of facilities technique, Share Program Library,
1965, SDA3391.
232. H.D. Sherali, The quadratic assignment problem: exact and heuristic algorithms, Ph.D.
thesis, Georgia Institute of Technology, Atlanta, June 1979.
233. H.D. Sherali and P. Rajgopal, A exible polynomial time construction and improvement
heuristic for the quadratic assignment problem, Computers & Operations Research 13
(1986), no. 5, 587{600.
234. B. Simeone, An asymptotically exact polynomial time algorithm for equipartition prob-
lems, Discrete Applied Math. 14 (1986), 283{293.
235. , Combinatorial optimization, Lecture Notes in Mathematics, vol. 1403, Springer-
Verlag, 1986.
236. C.C. Skiscim and B.L. Golden, Optimization by simulated annealing: A preliminary
computational study for the tsp, Proceedings of the 1983 Winter Simulation Conference,
1983.
237. J. Skorin-Kapov, Extensions of tabu search adaptation to the quadratic assignment prob-
lem, Computers and Operations Research forthcoming.
238. , Tabu search applied to the quadratic assignment problem, ORSA Journal on
Computing 2 (1990), no. 1, 33{45.
239. T.H.C. Smith, A computational comparision of an improved pair assignment algorithm
and a pair exchange algorithm for the quadratic assignment problem, Man. Sci. Res. Rep.
383, Carnegie Mellon University, Pittsburgh, Pa., November 1975.
240. L. Steinberg, The backboard wiring problem: A placement algorithm, SIAM Review 3
(1961), 37{50.
241. R.J. Stern and H. Wolkowicz, Inde nite trust region subproblems and nonsymmetric
eigenvalue perturbations, Tech. Report CORR 92-38, University of Waterloo, Waterloo,
Canada, 1992, to appear in SIOPT.
42 P.M. PARDALOS, F. RENDL, AND H. WOLKOWICZ
242. E. Taillard, Robust tabu search for the quadratic assignment problem, Parallel Computing
17 (1991), 443{455.
243. Y. Toyoda, A simpli ed algorithm for obtaining approximate solutions to zero-one integer
programming, Management Science 21 (1975), 1417{1427.
244. J. Turner and W.H. Kautz, A survey of progress in graph theory in the Soviet Union,
SIAM 12 (1970).
245. I. Ugi, J. Bauer, J. Friedrich, J. Gasteiger, C. Jochum, and W. Schubert, Neue anwen-
dungsgebiete fur computer in der chemie, Angew. Chemie 91 (1979), 99{184.
246. A. Vannelli and G. S. Rowan, An eigenvector based approach for ecient multi-stack
layout in vlsi design, 29th Midwest Symposium on Circuits and Systems (M. Ismail, ed.),
Noth Holland, New York, 1986, pp. 435{439.
247. M. Vlach, Branch and bound method for the three index assignment problem, Ekonomicko-
Matematicky Obzor (1967), 181{191.
248. T. E. Vollman and E. S. Bu a, The facilities layout problem in perspective, Management
Science 12 (1966), 450{468.
249. M. Werman, The relationship between integer and real solutions of constrained convex
programming, Mathematical Programming 51 (1991), 133{135.
250. D.H. West, Algorithm 608: Approximate solution of the quadratic assignment problem,
ACM Transactions on Mathematical Software 9 (1983), 461{466.
251. D.J. White, A parametric-based heuristic program for the quadratic assignment problem,
Naval Research Logistics Quaterly 40 (1993), no. 4, 553{568.
252. B. Whitehead and M. Z. Elders, An approach to the optimum layout of single-story
buildings, Architect's Journal 139 (1964), 1373{1380.
253. M.R. Wilhelm and T.L. Ward, Solving quadratic assignment problems by simulated an-
nealing, IEEE Transactions 19 (1987), no. 1, 107{119.
254. W.S. Wong and R.J.T. Morris, A new appraoch to choosing initial points in local search,
Information Processing Letters 30 (1991), 67{72.
(P.M. Pardalos) Department of Industrial and Systems Engineering, University of
Florida, Gainesville, FL 32611 USA and Technical University of Crete, Greece
E-mail address, P.M. Pardalos: pardalos@math.u .edu
(F. Rendl) Technische Universitat Graz, Institut fur Mathematik, Kopernikus-
gassa 24, A-8010 Graz, Austria
E-mail address, F. Rendl: rendl@ftug.dnet.tu-graz.ac.at
(H. Wolkowicz) University of Waterloo, Department of Combinatorics and Opti-
mization, Waterloo, Ontario, Canada
E-mail address, H. Wolkowicz: hwolkowicz@orion.uwaterloo.ca