See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/227094185

Integer programming formulations for the minimum weighted maximal

matching problem

Article in Optimization Letters · August 2011

DOI: 10.1007/s11590-011-0351-x

CITATIONS READS

13 1,204

2 authors:

Z. Caner Taşkın Tinaz Ekim

Bogazici University Bogazici University
55 PUBLICATIONS 672 CITATIONS 88 PUBLICATIONS 588 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Tinaz Ekim on 06 September 2016.

The user has requested enhancement of the downloaded file.

Optimization Letters manuscript No.
(will be inserted by the editor)

Integer Programming Formulations for the Minimum

Weighted Maximal Matching Problem

Z. Caner Taşkın · Tınaz Ekim

Received: date / Accepted: date

Abstract Given an undirected graph, the problem of finding a maximal

matching that has minimum total weight is NP-hard. This problem has been
studied extensively from a graph theoretical point of view. Most of the exist-
ing literature considers the problem in some restricted classes of graphs and
give polynomial time exact or approximation algorithms. On the contrary, we
consider the problem on general graphs and approach it from an optimization
point of view. In this paper, we develop integer programming formulations for
the minimum weighted maximal matching problem and analyze their efficacy
on randomly generated graphs. We also compare solutions found by a greedy
approximation algorithm, which is based on the literature, against optimal so-
lutions. Our results show that our integer programming formulations are able
to solve medium size instances to optimality and suggest further research for
improvement.

Keywords minimum maximal matching · edge dominating set · integer

programming

Z. Caner Taşkın’s research was partially supported by Boğaziçi University Research Fund
Grant No: 5028 and the research of Tınaz Ekim was supported by Grant 108M616 of CNRS-
TÜBİTAK joint research project, whose support is greatly acknowledged.

Z. C. Taşkın
Department of Industrial Engineering, Boğaziçi University
34342 Bebek, İstanbul, Turkey
E-mail: caner.taskin@boun.edu.tr
T. Ekim
Department of Industrial Engineering, Boğaziçi University
34342 Bebek, İstanbul, Turkey
E-mail: tinaz.ekim@boun.edu.tr
2

1 Introduction and Literature Survey

Given a graph, a matching is a set of edges that are pairwise non-adjacent.

A matching is said to be maximal if no other edge can be added to it while
keeping the property of being a matching. We say that an edge in a match-
ing dominates itself and all edges adjacent to it. Furthermore, nodes adjacent
to edges of a matching are said to be saturated by this matching. Note that
in a maximal matching of a graph G = (N, E) every edge in E is necessar-
ily dominated. The problem of finding a maximal matching having minimum
cardinality is called Minimum Maximal Matching (MMM). In the presence of
edge weights, the objective becomes minimization of the sum of the weights
in a maximal matching, and the related problem is called Minimum Weighted
Maximal Matching (MWMM).
MMM has been studied extensively due to its various applications. Yan-
nakakis and Gavril [16] describe an application related to a telephone switching
network built to route phone calls from incoming lines to outgoing trunks. The
worst case behavior of such a network (minimum number of calls routed when
the system is saturated) can be evaluated by a minimum maximal matching.
Another application appears in the theory of stable marriages; MMM gives an
upper bound on the number of unmatched candidates in a stable marriage,
which is naturally desired to be as small as possible [6].
A similar problem, called minimum edge dominating set (EDS), can be
stated as follows: given a graph G = (N, E), find a set of edges having min-
imum cardinality that dominates all edges in E. It is well known that the
size of a minimum maximal matching is equal to the size of a minimum edge
dominating set, and given a minimum edge dominating set, it is easy to find
a minimum maximal matching [16]. However, as pointed out in [8], this close
relationship is lost if edge weights are introduced.
Unlike the problem of finding a maximum matching in a given graph,
which is known to be polynomially solvable by Edmond’s augmenting path
algorithm [7], MMM is NP-hard in several classes of graphs including bipar-
tite or planar graphs with maximum degree 3 [16], planar bipartite graphs,
planar cubic graphs [9] and k-regular bipartite graphs for any fixed k ≥ 3
[6]. Similarly, while the maximum flow problem is polynomially solvable, the
minimum maximal flow problem is NP-hard [14]. In the literature, MMM is
handled mainly with graph theoretical tools. For instance, polynomially solv-
able cases for MMM can be found in [11] for trees, in [10] for block graphs,
in [12] for series-parallel graphs, and in [15] for bipartite permutation graphs
and cotrianglated graphs.
There are also some results on MMM from an approximation point of view.
It is well known that in general, a maximum matching cannot be larger than
twice the size of a minimum maximal matching since one edge in a minimum
maximal matching can cover at most two edges of a maximum matching.
Hence, any maximal matching gives a 2-approximation for MMM. Surprisingly,
no better approximation is known in general. Moreover, a 2-approximation can
also be obtained for the weighted version of EDS [8]. In [5], it is shown that
 3

MMM cannot be approximated within factor 67 unless P=NP, which leaves

space for a possible improvement between 76 and 2. More recently, Cardinal et
al. [4] give an approximation ratio strictly less than 2 for dense graphs. Slight
improvements on these results can be obtained by making certain assumptions
on the minimum degree and the density of the graphs ([3] and [5]).
To the best of our knowledge, MMM and MWMM have not been addressed
from an optimization point of view in the literature. Integer programming for-
mulations are used to derive approximation algorithms for weighted EDS [8]
and for a generalized version of weighted EDS, where capacities and demands
are introduced on edges [1]. However, these methods are not applicable to
MWMM; as already mentioned, the presence of weights breaks the close rela-
tionship between the two problems [8].
In this paper, we first give a natural integer programming formulation of
MWMM (Model 1). We then improve this formulation by adding new variables
to indicate saturated nodes in a solution (Model 2), which we then enhance by
re-formulating some constraints. This new formulation (Model 3) also allows
the derivation of some valid inequalities. We evaluate computational efficacy
of Model 1 and Model 3 on randomly generated graphs for both weighted and
unweighted edges. We also compare the results obtained by a greedy algorithm
for MWMM with optimal solutions. Our experiments show that Model 3 pro-
vides satisfactory results solving almost all instances with up to 170 nodes to
optimality in a reasonable amount of computational time. Furthermore, the
greedy algorithm gives good quality results for unweighted instances whereas
the quality considerably deteriorates when weights are introduced.

2 An Integer Programming Approach

Let G = (N, E) be an undirected graph and cij denote the weight associated
with edge (i, j) ∈ E. We define binary variable xij = 1 if edge (i, j) ∈ E is
selected in an optimal solution, and 0 otherwise. Let A(i) denote the set of
nodes that are adjacent to node i ∈ N . MWMM can be formulated as:
X
(Model 1) Minimize cij xij (1)
(i,j)∈E
X
subject to: xij ≤ 1 ∀i ∈ N (2)
j∈A(i)
X X
xik + xkj + xij ≥ 1 ∀(i, j) ∈ E (3)
k∈A(i) k∈A(j)
k6=j k6=i

xij ∈ {0, 1} ∀(i, j) ∈ E. (4)

The objective function (1) minimizes the total weight corresponding to the
selected edges. Constraints (2) enforce that at most one edge emanating from
each node can be selected (hence guaranteeing that the solution is a match-
ing). Constraints (3) ensure that each edge is either selected or is adjacent to
4

a selected edge, which guarantees that the generated matching is maximal.

Finally, Constraints (4) enforce binary restrictions on the x-variables.

3 An Improved Formulation

Our preliminary computational tests showed that Model 1 can only solve prob-
lems on small graphs to optimality due to the weak lower bound induced by the
linear programming relaxation (see Tables 1 and 2). In our tests, we observed
that integer programming solvers failed to generate effective cuts to improve
the lower bound. Furthermore, no special structure that can be exploited to
generate valid inequalities is apparent in Model 1. In order to alleviate these
P binary variable yi = 1 if node i ∈ N is saturated in
difficulties, let us define
the matching, that is j∈A(i) xij = 1. Using these additional variables, we can
formulate the problem as:

X
(Model 2) Minimize cij xij (5)
(i,j)∈E
X
subject to: xij ≤ 1 ∀i ∈ N (6)
j∈A(i)

yi + yj ≥ 1 ∀(i, j) ∈ E (7)
xij ≤ yi ∀i ∈ N, j ∈ A(i) (8)
X
yi ≤ xij ∀i ∈ N (9)
j∈A(i)

xij ∈ {0, 1} ∀(i, j) ∈ E (10)

0 ≤ yi ≤ 1 ∀i ∈ N. (11)

Similar to Model 1, the objective function (5) minimizes the total weight cor-
responding to the selected edges and Constraints (6) ensure that the generated
solution is a matching. Constraints (7) enforce the condition that at least one
end point of each edge has to be saturated in a maximal matching (otherwise
that edge can be added to the matching, contradicting maximality). Con-
straints (8) ensure that node i is saturated (yi = 1) if some edge emanating
from it is selected. Constraints (9) enforce the reverse condition that if node
i is saturated, then at least one of the edges emanating from it has to be
selected. Note that if the x-variables are binary-valued, then (8) and (9) will
force y-variables to take on binary values. Therefore, we relax y-variables as
continuous in (11).
We observe that Constraints (7) can be improved as follows: if xij = 1 for
some (i, j) ∈ E, then Constraints (8) imply that yi = yj = 1. Therefore, (7)
can be tightened Pas yi + yj − xij ≥ 1. Furthermore, Constraints (6), (8) and
(9) imply that j∈A(i) xij = yi in an optimal solution. Hence, the problem
 5

can equivalently be formulated as:

X
(Model 3) Minimize cij xij (12)
(i,j)∈E
X
subject to: xij = yi ∀i ∈ N (13)
j∈A(i)

yi + yj − xij ≥ 1 ∀(i, j) ∈ E (14)

xij ∈ {0, 1} ∀(i, j) ∈ E (15)
0 ≤ yi ≤ 1 ∀i ∈ N. (16)

Proposition 1 Model 1 is equivalent to Model 3.

ProofPLet x̂ be a solution of Model 1 having objective function value ẑ. Let

ŷi = j∈A(i) x̂ij , ∀i ∈ N . Then (x̂, ŷ) is a solution of Model 3 having the same
objective function value ẑ. Similarly, if (x̄, ȳ) is a solution of Model 3 having
objective function value z̄, then x̄ is a feasible solution of Model 1 having the
same objective function value z̄.

Note that Proposition 1 also holds for the linear programming relaxations
of the two models. Therefore, the linear programming bounds of the two mod-
els are equal. Both models have the same number of constraints and binary
variables. Model 3 has |N | additional (continuous) variables, which increases
the total number of variables but also makes the technology matrix more
sparse. Furthermore, Constraints (7) and (14) in our revised formulations re-
veal an embedded vertex cover structure in terms of the y-variables. Therefore,
valid inequalities exploiting the vertex cover structure can be devised. Berger
et al. [1] observe a similar structure in their formulation of the edge dominat-
ing set problem, and propose adding the well-known odd hole inequalities. Let
H ⊆ N denote the set of nodes that form an odd hole, that is the set of nodes
in H form a cycle having length equal to an odd number. Then, Constraints
(17) are valid [1]:
 
X |H|
yi ≥ ∀H ⊆ N and odd hole. (17)
2
i∈H

Also let C ⊆ N denote a clique in N . Since C has an edge between each node
pair that has to be covered as enforced by (14), at most one node in C can be
unsaturated (see also [2]) . Therefore, the following is valid:
X
yi ≥ |C| − 1 ∀C ⊆ N and clique. (18)
i∈C

These clique inequalities, although quite natural, do not seem to have been
used in integer programming formulations of the vertex cover problem in the
literature.
Finally, let i ∈ N be a node having degree 1, and j ∈ N be its only
neighbor. Since at least one of nodes i and j has to be saturated in a maximal
6

matching due to (14), and since the saturation of node i implies that node j is
also saturated by the matching, it follows that node j has to be saturated in
any maximal matching. Therefore, the following variable fixing rule is valid:

yj = 1 ∀(i, j) ∈ E, |A(i)| = 1. (19)

Remark 1 The valid inequalities (17)–(19) can also be expressed in terms of the
x-variables so that they can be used to tighten Model 1. Using the relationship
between the x- and y-variables enforced by (13), the odd hole inequalities (17)
can be written as follows
 
X X X |H|
2 xij + xij + xij ≥ ∀H ⊆ N and odd hole.
2
(i,j)∈E (i,j)∈E (i,j)∈E
i∈H,j∈H i∈H,j ∈H
/ i∈H,j∈H
/
(20)

Similarly, the clique inequalities (18) can be written as

X X X
2 xij + xij + xij ≥ |C| − 1 ∀C ⊆ N and clique, (21)
(i,j)∈E (i,j)∈E (i,j)∈E
i∈C,j∈C i∈C,j ∈C
/ i∈C,j∈C
/

and our fixing rule (19) can be expressed as

X
xi0 j = 1 ∀(i, j) ∈ E, |A(i)| = 1. (22)
i0 ∈A(j)

Note that (22) implies the corresponding (2) for each node that it is written for.
While (20)–(22) are equivalent to (17)–(19) with respect to their strengths, it
should be noted that they have a large number of nonzero coefficients, resulting
in significantly denser cuts.

4 Computational Results

We implemented the formulations discussed in the previous sections using

CPLEX 12.2 running on a Windows 7 PC with a 2.13 GHz Intel Core i3
CPU and 4 GB RAM. Our base test data set consists of randomly generated
problem instances for which the expected edge density of the graph (measured
2|E|
as D = |N |×(|N |−1) ) takes values 0.3, 0.5 and 0.7. In generating weighted
instances, we first generated random graphs as in the unweighted case, and
then assigned an integer weight uniformly distributed between 1 and 10 to
each edge. We generated ten problem instances for each problem size, which
is determined by the expected edge density and the number of nodes.
We used the default options of CPLEX for solving the integer program-
ming problems. Preliminary computational experience on Model 3 indicated
 7

that our variable fixing rule (19) and clique inequalities (18) are very effec-
tive at closing the integrality gap but use of the odd hole inequalities (17) is
not computationally justified. Therefore, in our experiments we only generated
(18) and (19). Specifically, we first enumerated all maximal cliques in the input
graph and generated a constraint of type (18) for each maximal clique identi-
fied. Note that there is an exponential number of maximal cliques in general;
however, for problem sizes that we tested they can be enumerated in a fraction
of Model 3’s solution time. We then added the generated constraints as “user
cuts,” which CPLEX checks for violation and automatically adds to tighten
the formulation as needed. We enforced (19) by setting the lower bound of the
corresponding yi -variable to 1. We imposed a 1200 seconds time limit (includ-
ing the cut generation time), past which we halted the execution of CPLEX in
all our experiments. Furthermore, we implemented Cardinal et al. [4]’s greedy
algorithm for MMM, which is known to have an approximation ratio of 2 (and
strictly less than 2 for dense graphs). While the algorithm in [4] is originally
given for the unweighted version of the problem, we slightly generalized it to
handle the weighted case. Our implementation of the greedy algorithm keeps
track of the edges in sets U and S corresponding to unprocessed and selected
edges, respectively:
– Step 0: U ← E, S ← ∅.
– Step 1: If U = ∅, then return S; else continue to Step 2.
– Step 2: Pick an edge (i, j) in U such that
cij
P
i0 j 0 ∈U ∩δ(i,j) ci j + cij
0 0

is minimized, where δ(i,j) is the set of all edges that are adjacent to (i, j).
– Step 3: Add (i, j) to S, remove (i, j) ∪ δ(i,j) from U , return to Step 1.
This algorithm behaves the same as the algorithm proposed in [4] for the
unweighted case, and also handles the weighted case in a similar manner.
However, the approximation analysis given in [4] explicitly assumes that all
edges have unit weight and therefore the same analysis does not hold for the
weighted case.
Our first experiment compares the performances of Model 1, Model 3 and
the greedy algorithm on the unweighted case. Our preliminary tests showed
that while the solvability of Model 1 is enhanced by the addition of valid
inequalities (21) and (22), the corresponding inequalities (18) and (19) are
significantly more effective in decreasing the overall solution time of Model
3. Therefore, we only added (18) and (19) to Model 3 in order to quantify
the effect of our enhancements over the basic formulation Model 1. Table 1
summarizes the results of Model 1, Model 3 and the greedy algorithm on graphs
having edge density D = 0.3, 0.5, and 0.7. For each problem size, we report
the following statistics calculated over ten random instances: (i) “Solved:” the
number of problem instances solved to optimality, (ii) “LP Gap:” the average
percentage gap between the linear programming relaxation and the optimal
objective function value, (iii) “Gap:” the average final percentage optimality
8

Table 1 Comparison of Model 1, Model 3 and greedy algorithm on small instances without
edge weights
Model 1 Model 3 Greedy
D N Solved LP Gap Gap Time Solved LP Gap Gap Time Gap Time
0.3 30 10 19.95 - 0.7 10 3.60 - 0.2 15.89 0.0
40 10 28.43 - 24.0 10 7.10 - 0.4 11.19 0.0
50 8 31.07 6.81 274.4 10 6.24 - 0.8 11.54 0.0
60 1 33.73 8.83 1111.0 10 6.41 - 1.9 10.58 0.0
70 0 35.00 10.02 - 10 6.32 - 2.9 10.25 0.1
0.5 30 10 31.14 - 48.4 10 4.92 - 0.3 9.87 0.0
40 2 35.17 7.96 723.1 10 4.54 - 0.5 5.69 0.0
50 0 37.83 10.72 - 10 5.50 - 1.2 6.68 0.0
60 0 38.59 11.67 - 10 4.36 - 2.8 8.73 0.1
70 0 40.52 13.97 - 10 5.54 - 5.2 7.37 0.1
0.7 30 10 34.47 - 53.2 10 2.41 - 0.3 8.46 0.0
40 0 39.17 11.56 - 10 4.12 - 0.6 4.97 0.0
50 0 40.56 12.36 - 10 4.23 - 1.8 6.80 0.0
60 0 41.50 12.93 - 10 3.57 - 3.2 6.07 0.1
70 0 42.71 14.38 - 10 4.21 - 5.2 5.16 0.2

gap for instances that could not be solved within the allowed time limit by the
formulations (for the greedy algorithm we report the average percent difference
between greedy algorithm’s result and the optimal objective function value
found by Model 3) , (iv) “Time:” the average amount of time in seconds spent
by each algorithm (for Model 1 and Model 3 we report the average on the
instances that were solved to optimality within the allowed time limit).
Out of the 150 instances in this data set, CPLEX could solve Model 1
to optimality for only 51 instances within 1200 seconds while being able to
solve all instances to optimality for Model 3 within a few seconds. The re-
sults show that Model 3 significantly outperforms Model 1. Note that linear
programming relaxations of Model 1 and Model 3 are equivalent (Proposition
1). Therefore, the difference between the “LP Gap” values of the two formu-
lations is the result of our clique inequalities and variable fixing rule, which
are highly effective in closing the optimality gap. We observe that the linear
programming relaxation gets weaker and our inequalities get more effective
as the density increases. The greedy algorithm executed very quickly on all
problem instances in this data set. While the greedy algorithm could find an
optimal solution for only 9 problem instances, it generated solutions having
an average optimality gap of 8.62%, which corresponds to choosing 1.88 more
edges than an optimal solution. We observe that the performances of Model 1
and Model 3 deteriorate as the number of nodes in the graph and edge density
increase. This is expected since the number of binary variables in both formu-
lations increase with the number of edges. On the other hand, note that the
quality of the solution found by the greedy algorithm tends to improve as the
problem size increases. This can be explained by observing that as the total
number of edges increases, the selection decision about each individual edge
becomes less critical in the overall solution quality.
 9

Table 2 Comparison of Model 1, Model 3 and greedy algorithm on small instances with
edge weights
Model 1 Model 3 Greedy
D N Solved LP Gap Gap Time Solved LP Gap Gap Time Gap Time
0.3 70 10 15.58 - 2.5 10 4.67 - 0.8 29.76 0.1
80 10 18.33 - 5.8 10 4.75 - 1.2 30.56 0.1
90 10 21.43 - 18.6 10 5.07 - 1.9 34.76 0.2
100 10 21.58 - 45.5 10 4.60 - 2.6 33.24 0.3
110 4 23.95 6.11 421.1 10 5.66 - 5.8 30.12 0.4
120 1 25.24 7.53 307.7 10 6.33 - 11.0 29.36 0.6
0.5 70 10 24.41 - 33.1 10 3.65 - 1.1 31.22 0.1
80 10 25.61 - 220.1 10 4.20 - 2.0 28.32 0.2
90 1 27.75 6.63 597.0 10 4.12 - 4.0 27.80 0.3
100 0 30.33 10.26 - 10 4.67 - 6.8 20.86 0.5
110 0 31.51 13.38 - 10 5.02 - 18.5 25.98 0.7
120 0 33.54 16.87 - 10 5.91 - 34.9 24.75 1.0
0.7 70 9 28.56 8.81 233.5 10 2.50 - 2.1 28.80 0.2
80 0 30.79 8.14 - 10 3.02 - 3.9 25.31 0.3
90 0 33.07 12.15 - 10 3.57 - 7.2 23.98 0.4
100 0 35.43 15.53 - 10 5.24 - 23.2 18.97 0.7
110 0 36.64 16.51 - 10 4.93 - 41.1 19.85 1.0
120 0 37.78 18.53 - 10 5.03 - 90.3 20.37 1.4

Our second experiment compares the three approaches on weighted prob-

lem instances. We observe by comparing Tables 1 and 2 that Model 1 is able
to solve larger problem instances to optimality compared to the unweighted
case. This result is not surprising since it is known that symmetry can make
integer programming problems very difficult to solve (see [13]) and adding edge
weights eliminates some symmetry by differentiating edges. As before, Model
3 is able to solve all problem instances in this data set within a few seconds
of CPU time. We also observe that the quality of the solutions generated by
the greedy algorithm is significantly lower than the results for the unweighted
case.
Our next set of experiments are aimed at analyzing the performances of
Model 3 and greedy algorithm on larger problem instances. Table 3 summa-
rizes the results of tests performed on unweighted graphs. As before, we report
the percentage optimality gaps of the linear programming relaxation with and
without our clique inequalities and variable fixing rule. We note that observa-
tions similar to Table 1 can be made about the performance of Model 3. In
particular, we note that the problem becomes harder to solve for CPLEX as D
and N increase, which results in a higher number of binary variables in Model
3. Comparison of the optimality gaps calculated for the greedy algorithm on
this data set with Table 1 shows that the greedy algorithm tends to get closer
to optimality as N and D increase. Note that optimal solutions of some prob-
lem instances having N ≥ 170 are not known. Since we used the best lower
bound reported by CPLEX for calculating greedy algorithm’s gaps, the gaps
reported for these instances possibly overestimate the true gap between the
greedy algorithm’s solution and the optimal objective function value.
10

Table 3 Comparison of Model 3 and greedy algorithm on large instances without edge
weights
Model 3 Greedy
D N Solved LP Gap (No cuts) LP Gap Gap Time Gap Time
0.3 150 10 42.40 9.34 - 197.6 6.51 1.4
160 10 42.94 9.67 - 374.8 5.58 1.8
170 10 43.03 9.45 - 743.4 5.61 2.3
180 4 43.54 9.81 2.78 938.8 7.31 2.9
190 0 44.05 10.21 3.99 - 8.66 3.6
0.5 150 10 45.09 7.28 - 255.8 4.03 2.5
160 10 45.05 6.86 - 353.3 3.67 3.2
170 7 45.51 7.33 3.22 753.5 4.84 4.1
180 1 46.04 7.91 3.08 1191.7 5.79 5.2
190 0 46.26 8.06 4.51 - 7.72 6.5
0.7 150 10 46.14 5.34 - 558.1 3.30 3.5
160 10 46.33 5.46 - 901.4 3.84 4.6
170 2 46.69 5.82 3.57 1024.8 5.68 5.9
180 0 46.82 5.82 5.23 - 7.71 7.4
190 0 47.06 6.01 5.83 - 8.27 9.2

Table 4 Comparison of Model 3 and greedy algorithm on large instances with edge weights
Model 3 Greedy
D N Solved LP Gap (No cuts) LP Gap Gap Time Gap Time
0.3 150 10 28.57 7.07 - 56.4 28.41 1.4
160 10 30.24 7.68 - 160.0 27.67 1.8
170 10 31.16 7.78 - 186.6 26.88 2.4
180 8 33.04 8.73 2.81 501.2 29.33 3.0
190 5 33.71 9.29 4.36 650.1 27.19 3.7
0.5 150 10 36.79 6.87 - 129.6 20.90 2.5
160 10 37.59 6.83 - 273.3 20.07 3.3
170 10 38.67 7.17 - 457.0 17.93 4.2
180 5 39.35 7.36 2.40 785.9 17.69 5.4
190 2 40.17 7.77 4.04 1182.4 20.10 6.7
0.7 150 10 40.33 5.30 - 340.1 15.61 3.6
160 10 41.48 5.45 - 472.0 12.45 4.7
170 9 41.93 5.68 3.83 860.7 14.87 5.9
180 3 42.52 5.71 2.57 1023.0 14.07 7.5
190 0 43.16 5.90 4.53 - 13.90 9.4

Finally, Table 4 summarizes the results obtained on large instances having

weighted edges by Model 3 and the greedy algorithm. As before, Model 3
can solve more weighted problem instances to optimality than unweighted
instances (112 and 84, respectively). However, unlike Model 1, we observe
that the difference between solvable problem sizes is not very significant for
Model 3. This can be explained by the fact that our clique inequalities (18)
and variable fixing rule (19) are solely based on the topology of the graph, and
they do not use any information about edge weights. Comparison of Tables 2
and 4 reveals that quality of the solutions generated by the greedy algorithm
tends to improve as N and D increase, which is in alignment with our earlier
observations.
 11

5 Conclusions and Future Research

To the best of our knowledge, this paper presents the first integer program-
ming formulations and computational results on MMM and MWMM. Our
results constitute a basis for the comparison of the performances of various
exact methods, heuristics and approximation algorithms. One potential fu-
ture research topic is to adjust our formulations to specific graph types. Even
though our formulations can be used on general graphs, it may be possible
to improve them by deriving bounds and valid inequalities for specific graph
types. A particular graph type of interest in the literature is regular graphs,
i.e. graphs where all nodes have the same degree. It is known that MMM is
NP-hard in regular bipartite graphs but performances of various heuristics and
integer programming formulations such as Models 1 and 3 on regular graphs
are unknown. More studies are needed to answer whether the structure of
regularity makes it easier to solve MWMM, or the high symmetry caused by
regularity makes the formulations and heuristics less effective.

References

1. A. Berger, T. Fukunaga, H. Nagamochi, and O. Parekh. Approximability of the ca-

pacitated b-edge dominating set problem. Theoretical Computer Science, 385:202–213,
2007.
2. I. M. Bomze, M. Budinich, P. M. Pardalos, and M. Pelillo. The maximum clique problem.
In D.-Z. Du and P. M. Pardalos, editors, Handbook of Combinatorial Optimization.
Kluwer Academic Publishers, 1999.
3. J. Cardinal, M. Labbé, S. Langerman, E. Levy, and H. Mélot. A tight analysis of the
maximal matching heuristic. COCOON, Lecture Notes in Computer Science, 3595:701–
709, 2005.
4. J. Cardinal, S. Langerman, and E. Levy. Improved approximation bounds for edge
dominating set in dense graphs. Theoretical Computer Science, 410:949–957, 2009.
5. M. Chlebı́k and J. Chlebı́ková. Approximation hardness of edge dominating set prob-
lems. Journal of Combinatorial Optimization, 11(3):279–290, 2006.
6. M. Demange and T. Ekim. Minimum maximal matching is NP-hard in regular bipartite
graphs. TAMC 2008, Lecture Notes in Computer Science, 4978:364–374, 2008.
7. J. Edmonds. Paths, trees and flowers. Canadian Journal of Mathematics, 17:449–467,
1965.
8. T. Fujito and H. Nagamochi. A 2-approximation algorithm for the minimum weight
edge dominating set problem. Discrete Applied Mathematics, 118:199–207, 2002.
9. J.D. Horton and K. Kilakos. Minimum edge dominating sets. SIAM Journal on Discrete
Mathematics, 6(3):375–387, 1993.
10. S.F. Hwang and G.J. Chang. The edge domination problem. Discuss. Math. Graph.
Theory, 15(1):51–57, 1995.
11. S.L. Mitchell and S.T. Hedetniemi. Edge domination in trees. In Proceedings of the
8th Southeastern Conference on Combinatorics, Graph Theory and Computing, pages
489–509, Louisiana State University, Baton Rouge, La., 1977.
12. M.B. Richey and R.G. Parker. Minimum-maximal matching in series-parallel graphs.
European Journal of Operations Research, 33(1):98–105, 1988.
13. H. D. Sherali and J. C. Smith. Improving discrete model representations via symmetry
considerations. Management Science, 47:1396–1407, 2001.
14. J. Shi and Y. Yamamoto. A global optimization method for minimum maximal flow
problem. ACTA Mathematica Vietnamica, 22(1):271–287, 1997.
 12

15. A. Srinivasan, K. Madhukar, P. Nagavamsi, C. Pandu Rangan, and M.-S. Chang. Edge
domination on bipartite permutation graphs and cotriangulated graphs. Information
Processing Letters, 56(3):165–171, 1995.
16. M. Yannakakis and F. Gavril. Edge dominating sets in graphs. SIAM Journal on
Applied Mathematics, 38:364–372, 1980.

View publication stats