SOME PRIMAL-DUAL THEOREMS IN

GRAPH THEORY - A STUDY

A THESIS
Submitted by
VENKATA HARISH IMMADI
In partial fulllment for the award of the degree of
MASTER OF TECHNOLOGY
IN
COMPUTER SCIENCE AND ENGINEERING (INFORMATION
SECURITY)
Under the guidance of
Dr. K. Murali Krishnan
DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING
NATIONAL INSTITUTE OF TECHNOLOGY CALICUT
NIT CAMPUS PO, CALICUT
KERALA, INDIA 673601
May, 2012
ii
ACKNOWLEDGEMENTS
First of all, I would like to express my gratitude to my guide Dr.K Murali
Krishnan, Department of Computer science and Engineering, National Institute
of Technology, Calicut for his inspiring guidance, motivation, support and encour-
agement. His strenuous eorts and valuable pieces of advice were available to me
throughout my endeavor without which it would have been impossible for me to
complete this work in time.
I would like to express my sincere thanks to the project coordinator Mr. G
Gopa Kumar for allowing me to do this project.
I take this opportunity to express my sincere thanks to, Dr.M. N. Bandy-
opadhyay, Director, National Institute of Technology, Calicut for providing me
with the facilities to carry out the work.
I am thankful to the teaching and non teaching sta for their support and
encouragement throughout my work.
I Venkata Harish
DECLARATION
I hereby declare that this submission is my own work and that, to the best of my
knowledge and belief, it contains no material previously published or written by
another person nor material which has been accepted for the award of any other
degree or diploma of the university or other institute of higher learning, except
where due acknowledgment has been made in the text.
Place: Calicut
Date:
Signature :
Name : I Venkata Harish
Reg.No: M100170CS
CERTIFICATE
This is to certify that the thesis entitled: SOME PRIMAL-DUAL THEO-
REMS IN GRAPH THEORY - A STUDY submitted by Sri VENKATA
HARISH IMMADI to National Institute of Technology Calicut towards partial
fulllment of the requirements for the award of Degree of Master of Technology in
Computer Science Engineering is a bona de record of the work carried out by him
under my supervision and guidance.
Signed by Thesis Supervisor with name and date
Place: Calicut
Date:
Signature of Head of the Department
Oce Seal
v
Contents
Chapter
1 Introduction 1
1.1 The Max-ow Min-cut Theorem . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.3 Proof of the Max-ow Min-cut Theorem . . . . . . . . . . . 6
2 Proving Graph Theorems using Max-ow Min-cut Theorem 9
2.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Mengers Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Konig-Egervary Therorem . . . . . . . . . . . . . . . . . . . . . . . 14
3 Max-ow Min-cut Theorem and Konigs Theorem using Total Unimodular-
ity 16
3.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Konigs Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3 The Max-Flow Min-Cut Theorem . . . . . . . . . . . . . . . . . . . 24
4 The Concurrent Multi-commodity Flow Problem 31
4.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.2 Linear Programming Formulation . . . . . . . . . . . . . . . . . . . 32
vi
5 Conclusion 39
Bibliography 41
vii
Abstract
The objective of this report is to present some central min-max theorems
in graph theory from the view point of linear programming. In the introduc-
tory chapter, we review the well known Max-ow Min-cut theorem and discuss
the standard proof using Ford-Fulkerson algorithm. The next chapter will discuss
some graph theoretic consequences of the Max-ow Min-cut theorem in combi-
natorial optimization. In the third chapter we present the proof of the Max-ow
Min-cut theorem and Konigs theorem using the properties of total unimodular
matrices in linear programming. In the fourth chapter, we discuss the problem
of Concurrent Multi-commodity Flow(CMFP) and present a linear programming
formulation.
viii
Figures
Figure
1.1 (a) A network with edge capacities. (b) The corresponding ow network
with |f| = 4. The dotted line forms a cut which is also a minimum cut . 3
1.2 (a) A network N of positive integral edge capacities. (b) The residual
graph obtained after pushing a ow of 1 through the path s c b a t 5
1.3 (a) The initial graph(network) G. (b) The residual graph G

after sending
a ow of 1 unit along the path s c d t. (c) The nal residual graph
G

after sending a ow of 1 unit along the path s a d c b t. . . 8

2.1 (a) The paths P
1
and P
2
where (a, b) P
1
, (b, a) P
2
. (b) The corre-
sponding paths P
3
and P
4
. . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 (a) A bipartite graph with partitions {a, b, c} and {d, e, f}. (b) The
resulting network after transformation. . . . . . . . . . . . . . . . . . . 15
4.1 An example of a two commodity ow network with unit demands on the
commodities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Chapter 1
Introduction
In this report, we study some min-max theorems in combinatorial optimiza-
tion. One of the earliest and well recognised result is the Max-ow Min-cut theorem
by Ford and Fulkerson [5] in 1956. This theorem states the equivalence between
the maximum ow and minimum cut in a network. In this chapter we review the
Max-ow Min-cut theorem and its proof using Ford-Fulkerson algorithm. This
theorem had several implications in combinatorial optimization, several other the-
orems can be viewed as a consequence of this theorem [11]. In second chapter,
we presen Konigs and Mengers theorems as consequences of Max-ow Min-cut
theorem. The Mengers theorem provide bounds on the connectivity of a graph.
There are two versions of Mengers theorem [1], edge version and vertex version.
This theorem is a special case of the Maximum ow problem and can be proved by
using the Max-ow Min-cut theorem. Konigs theorem [13, 9] states the equality
between maximum matching and minimum vertex cover in bipartite graph.
All the above theorems are min-max theorems, that is, if one of the problem is
a maximization problem then the other is a minimization problem. These problems
can be encoded as linear programs. A Chandra Babu et al.[2] has given new proofs
for Max-ow Min-cut theorem and Konigs theorem using total unimodularity of
the coecient matrix in their linear programming formulations. The third chapter
is a study of the proofs given by A Chandra Babu et al.[2]. We have lled a few
missig links while presenting this material.
2
The Multi-commodity ow problem is a more generalized network ow prob-
lem. The problem of nding a maximum ow in a multi-commodity network arises
in many network instances. In a Multi-commodity ow problem, there exists k 1
commodities each having its own source and sink. In the fourth chapter, we give
an introduction to the Concurrent Multi-commodity Flow problem(CMFP) [15]
1
and present the linear programming formulation for the problem and its dual.
In CMFP, every commodity is assigned a demand D
i
, our objective is to assign
ows to the commodities so as to maximize a fraction such that the ow for any
commodity is at least D
i
units.
1.1 The Max-ow Min-cut Theorem
1.1.1 Preliminaries
Denition 1.1.1. (Network) A Network is a directed graph G(V, E, c, s, t) with
vertex set V and an arc set E in which every directed edge (i, j) E, has a non
negative capacity c(i, j) 0, c : V V R
+
. The vertices s V, t V are the
source vertex and the sink vertex of the network.
Denition 1.1.2. (s,t-cut) An s, t-cut is a partitioning of the vertices of a Network
into two sets, say A and B, such that s A and t B.
The capacity of a cut (A, B) is given as
C(A, B) =

uA,vB
c(u, v)
Denition 1.1.3. (Flow) A ow is a mapping f : E R, denoted by f
uv
or
f(u, v), subject to the following constraints:
(1) f(u, v) c(u, v) for each (u, v) E (capacity constraint)
(2) f(u, v) = f(v, u) (skew symmetry)
1
It is also referred to as The Maximum Concurrent Flow Problem .
3
(3)

vV
f(u, v) = 0 u V \ {s, t} (conservation of ow).
The value of ow is dened by |f| =

vV
f(s, v) , where s is the source of
N. It represents the amount of ow passing out of the source to the sink. Figure
1.1 shows a typical network and a ow in that network. Here f can be viewed as
a vector over the directed edges
2
of the network.
1
4
2
5
1
1
4
2
s
t
a
b
c
d
(a)
1/1
1/1
1/1
1/2
2/2
2/4
2/5
3/4
s
t
a
b
c
d
(b)
Figure 1.1: (a) A network with edge capacities. (b) The corresponding ow network
with |f| = 4. The dotted line forms a cut which is also a minimum cut
Denition 1.1.4. (Flow across a cut) If (A, B) is any cut in the network, then
the ow across the cut is dened as
f(A, B) =

uA,vB
f(u, v)
Lemma 1.1.1. If f is a ow in a network G, then for any s, t-cut (A, B), f(A, B) =
|f|.
Proof. We can prove this by considering the ow over the cut and reducing it to
the ow from the source vertex which is the actual ow.
2
The terms directed edge and arc may be used interchangeably.
4
f(A, B) =

uA,vB
f(u, v)
=

uA,vB
f(u, v) +

{w,w

}A
f(w, w

)
Since,

{w,w

}A
f(w, w

) = 0. Because for every {x, y} A both f(x, y) and

f(y, x) are added.
=

uA,vV
f(u, v)
=

vV
f(s, v) +

uA\{s},vV
f(u, v)
=

vV
f(s, v)
= |f|
Denition 1.1.5. (Residual Graph) If f is a ow in a graph G, then the residual
graph is dened as G
f
with c
f
(u, v) = c(u, v) f(u, v).
For a directed edge (a, b) E, if c(a, b) = 0 then in G
f
, the residual capacity
will be c
f
(a, b) = 0 f(a, b) = f(a, b). Figure 1.2 gives an example for a residual
graph(network).
Lemma 1.1.2. f

is a ow in G
f
i f +f

is a ow in G.
Proof. We prove this by proving the following two conditions.
(1) A ow f

is a ow in G
f
i it satises f

(u, v) c
f
(u, v).
f

(u, v) c
f
(u, v)
f

(u, v) c(u, v) f(u, v)

f(u, v) +f

(u, v) c(u, v)
f +f

(u, v) c(u, v)
5
s
b
t
a
c
1
2
2
1
3
3
(a)
s
b
t
a
c
0
0
2
3
1
1
1
1
1
2
(b)
Figure 1.2: (a) A network N of positive integral edge capacities. (b) The residual graph
obtained after pushing a ow of 1 through the path s c b a t
(2) |f f

| = |f| |f

|
|f f

| =

vV
(f f

)(s, v)
=

vV
[f(s, v) f

(s, v)]
=

vV
f(s, v)

vV
f

(s, v)
= |f| |f

|
Lemma 1.1.3. If f

is maximum in G
f
, then f +f

is maximum in G.
Proof. Suppose if some ow |g| > |f + f

| in G, then it is easy to see that g f

is a ow in G
f
|g f| = |g| |f| > |f +f

| |f|
= |f| +|f

| |f| = |f

|
But, this leads to a contradiction to the fact that |f

| is maximum in G
f
.
6
Denition 1.1.6. (Capacity of a path) The capacity of a path P is given as,
C(P) = Min
(u,v)P
c(u, v)
Denition 1.1.7. (Augmenting Path) An augmenting path P is an s, t-path of
positive capacity.
3
1.1.2 Bounds
Let (A, B) be any s, t-cut, then the value of the ow is at most the capacity
of the cut.
|f| C(A, B)
Since, from Lemma 1.1.1
|f| = f(A, B)
=

uA,vB
f(u, v)

uA,vB
c(u, v) = C(A, B)
1.1.3 Proof of the Max-ow Min-cut Theorem
Theorem 1. (Ford-Fulkerson, 1956) [5] In a Network G, let f be any maximum
ow in G, then a cut (A, B) for which f(A, B) = c(A, B)
Proof. The proof given by Ford-Fulkerson is an algorithmic proof. Here, we give
the algorithm that iteratively nds an augmenting path and augments the ow
and computes the residual graph.
The total value of the resulting ow will be f
1
+ f
2
+ f
3
. . . + f
k
, where k
is the number of iterations and f
i
is the ow obtained in the i
th
iteration. The
above algorithm terminates by producing a maximum ow and also we can show
the existence of a cut which will be the minimum cut. The gure 1.3 shows the
execution of the Ford-Fulkerson algorithm on an example network.
3
Here path means a directed path from s to t.
7
Algorithm 1 Maxow(G)
Find an augmenting path P in G
if P does not exist then
return 0
else
let f be a ow of value C(P) along P
return (f +Maxflow(G
f
))
end if
In order to prove this theorem, we rst consider the following three state-
ments. And by showing a circular dependency between them, we prove the theo-
rem.
(1) G has a Maximum ow |f|
(2) G
f
has no augmenting path
(3) Some cut of G is saturated
As we can see, 1 2 and 3 1. These two are trivial. Because, if
there exists an augmenting path, the ow can be further increased by augmenting
through that path. But, this contradicts the fact that f is maximum. And, for the
second, since the ow can not exceed the capacity of a cut, if some cut is saturated,
then the ow has to be maximum.
What remains is to prove is that 2 3. For this, consider two sets of vertices
A and B in G
f
where A consists of all the vertices reachable from s and B consists
of all the vertices from which t is reachable. Note that the graph we consider is
the residual graph G
f
. Now, we can say that s / B and t / A. If any of the two
are not correct, it implies that there exists an augmenting path, which contradicts
the hypothesis. Now, consider the cut (A, B), this forms a saturated cut. Hence
proved.
8
s t
a b
c
d
(a)
1
1
1
1
1
1
1
(b)
s
1
a b
t
d c
1
0
0
1
1
1 1
1
0
t
(c)
b
c
d
0
0
1
0
1
1
0
0
1
a
s
1
0
1
1
0
Figure 1.3: (a) The initial graph(network) G. (b) The residual graph G

after sending
a ow of 1 unit along the path s c d t. (c) The nal residual graph G

after sending a ow of 1 unit along the path s a d c b t.

Chapter 2
Proving Graph Theorems using Max-ow Min-cut Theorem
In the previous chapter weve seen the proof for Max-ow Min-cut theorem.
In this chapter we will present other theorems in graph theory that can be viewed
as special cases of the Max-ow Min-cut theorem. The Mengers theorem [1]
characterizes the connectivity of a nite graph. There are two versions of it, the
edge version and the vertex version. Below, we present a proof of both the versions
using the Max-ow Min-cut theorem. The Konig-Egervary theorem [9] describes
an equivalence between the maximum matching problem and the minimum vertex
cover problem in a bipartite graph. A proof of this theorem is also presented using
the above Max-ow Min-cut theorem. The results or the proofs presented are not
unknown previously. The work presented here is only a consolidation of the known
facts scattered in literature to present complete and comprehensive proofs for the
graph theoretic min-max results discussed here.
2.1 Preliminaries
Denition 2.1.1. (Bipartite Graph) A graph G(V, E) in which there exists a par-
titioning of vertices into two sets X and Y such that (x, y) E x X, y Y
Denition 2.1.2. (Vertex cover) A vertex cover in a graph G(V, E) is a set S of
vertices such that (x, y) E either x S or y S.
10
Denition 2.1.3. (Minimum Vertex Cover) A vertex cover that contains the min-
imum number of vertices.
Denition 2.1.4. (Matching) A matching in a graph G(V, E) is a set M of edges
such that no two edges contain a common vertex.
Denition 2.1.5. (Maximum matching) A maximum matching is a matching that
contains the maximum number of edges.
A graph may contain more than one maximum matching and one minimum
vertex cover.
Denition 2.1.6. (Internally-disjoint paths) Two paths are internally-disjoint if
neither of them contains a non-endpoint vertex of the other.
Denition 2.1.7. (Edge-disjoint paths) Two paths are edge-disjoint if neither of
them contains a common edge.
Denition 2.1.8. (x,y-separator) Given x, y V (G), a set S V (G) {x, y} is
an x, y-separator if GS has no x, y-path.
Denition 2.1.9. (x,y-cut) Given x, y V (G), a set C E(G) is an x, y-cut if
GC has no x, y-path.
2.2 Mengers Theorem
In order to apply the Max-ow Min-cut theorem, we may need to transform
the given undirected graph G into a directed graph G

. This can be done by

replacing every undirected edge (a, b) in G with two directed edges (a, b) and
(b, a), one forward and backward, in G

.
Lemma 2.2.1. In a graph G(V, E), the number of pairwise edge-disjoint x, y-paths
in G

are equal to the number of pairwise edge-disjoint paths in G, where G

is the
corresponding directed graph.
11
Proof. It is easy to see that there will be at least as many pairwise edge-disjoint
x, y-paths in G

as there are in G. Because if two paths are edge disjoint in G,

they are edge-disjoint in G

also.
1
P
P
2
P
2
1
P
y
a
b
x
(a)
3
P
P
3
P
4
4
P
y
a
b
x
(b)
Figure 2.1: (a) The paths P
1
and P
2
where (a, b) P
1
, (b, a) P
2
. (b) The correspond-
ing paths P
3
and P
4
.
To show that there can be no more in G

, consider two edge-disjoint paths

P
1
, P
2
in G

. If there is no edge (a, b) such that (a, b) P

1
and (b, a) P
2
, then
the paths have to be edge-disjoint in G also. But, if such an edge exists, then we
can form two new paths that does not have either (a, b) or (b, a). A path P
3
can
be formed by joining the x, a-path in P
1
with a, y-path in P
2
and a path P
4
can
be formed by joining the x, b-path in P
2
with b, y-path in P
1
. Now, the paths P
3
and P
4
does not contain either (a, b) or (b, a). So, all such edges can be eliminated
and we can construct two paths corresponding to P
1
and P
2
that are edge-disjoint
in G

and these have to be edge-disjoint in G also. In this way all the pairwise
edge-disjoint x, y-paths in G

correspond to pairwise edge-disjoint x, y-paths in G.

Hence the lemma is proved.
Lemma 2.2.2. In a graph G(V, E), the number of pairwise internally-disjoint
x, y-paths in G

are equal to the number of pairwise internally-disjoint paths in G,

where G

is the corresponding directed graph.

12
Proof. Let k be the maximum number of pairwise internally-disjoint x, y-paths in
G and U be the set of these paths. Then the corresponding directed paths in G

will also be pairwise internally-disjoint. Let U

be the corresponding set of paths

in G

. Then there can not exist any more paths in G

that are pairwise internally-

disjoint to all the paths in U

. Suppose there exists such a path P in G

that is
internally-disjoint to every path in U

. Then it implies that P does not have a

vertex in common with any of the paths in U

. Then the corresponding undirected

path in G should also be pairwise internally-disjoint to all the paths in U. But,
this leads to a contradiction to the maximality of U in G. Hence |U

| = k.
Theorem 2. (Mengers,[1927])(Edge version) If x, y are two vertices in a graph
G and (x, y) / E(G), then the minimum size of an x, y-cut equals the maximum
number of pairwise edge-disjoint x,y-paths
Proof. The graph G(V, E) is an undirected graph. Let k be the maximum number
of pairwise edge-disjoint x, y-paths(undirected) in G. The necessary part of the
theorem can be easily veried. Since the minimum x, y-cut should contain at least
one edge from each x, y-path, it should be at least k. Therefore, the size of the
minimum x, y-cut is k.
In order to prove suciency, we rst transform the given undirected graph
into a network with x as source and y as sink. Let G

be the new directed

graph(Network). Assign unit capacities to all the directed edges in G

. Then
by lemma 2.2.1 the number of pairwise edge-disjoint x, y-paths in G

is equal to
k. Since the edges are of unit capacity, the maximum ow possible will be equal
to the number of pairwise edge-disjoint x, y-paths. So maximum ow in G

is k.
Now, by applying the Max-ow Min-cut theorem, we can see that the size of the
minimum cut(directed) in G

is k. It is easy to see that the corresponding set of

undirected edges in G also form a cut in G. And since every cut in G is at least
k, this will be minimum.
13
Theorem 3. (Vertex version) If x, y are two vertices in a graph G and (x, y) /
E(G), then the minimum size of an x, y-separator equals the maximum number of
pairwise internally-disjoint x, y-paths.
Proof. Let k be the maximum number of pairwise internally-disjoint x, y-paths in
G. Then the size of the minimum x, y-separator should be at least k since it should
include at least one vertex from each path.
In order to prove suciency, we will use the Max-ow Min-cut theorem.
First, we will transform the undirected graph G to a directed graph G

. From
lemma 2.2.2, the number of pairwise internally-disjoint paths in G

are k. Now,
we will transform the directed graph G

as follows. Let G

be the new graph. For

every vertex v G

, include two vertices in G

, v
in
and v
out
. v
in
consists of all
the inward edges to v and v
out
consists of all the outward edges from v. Add a
directed edge from v
in
to v
out
. Assign unit capacity to all the edges of G

. x
in
and
y
out
can be left out because there will be no ow into the source vertex and out of
sink vertex.
Now, we show that the maximum ow in G

is k. Clearly, every path in

G

correspond to unique path in G

. A path entering a vertex v

in
has to leave
through the directed edge (v
in
, v
out
) and this edge is of unit capacity. This implies,
no two augmenting paths in G

can have a common vertex. The k paths in G

that
correspond to the pairwise internally-disjoint x, y-paths in G

will each allow for a

ow of one unit. There can not be any more ow. If the ow is more than k, then
there exists an augmenting path that is pairwise internally-disjoint to all other
augmenting paths in G

. Then the corresponding path in G

will also be pairwise

internally-disjoint to all other paths in G

. But, this leads to a contradiction to

the fact that k is maximum in G

.
Now, by applying the Ford Fulkerson algorithm, we can get a cut of size k.
14
Select a set S of vertices for each edge in the cut as follows:
(1) If the edge is (v
in
, v
out
), select v.
(2) If the edge is (v
out
, u
in
) where {u, v} / {s, t}, select either u or v.
(3) In case of (x
out
, v
in
) or (v
out
, y
in
), select the vertex v.
Clearly, the size of S is k. This set also forms a separator in the network
G

. Otherwise, there exists at least one more x, y-path in G

from which no edge

is included in S. But, the corresponding path in the modied network G

will be
a x, y-augmenting path. This contradicts the fact that f is maximum in G

. Also
it is easy to see that this set S will be a separator in original graph G. And since
the size of the minimum separator is at least k, this is minimum. Hence proved.
2.3 Konig-Egervary Therorem
Theorem 4. (Konigs theorem,[1931]) In any bipartite graph, the number of edges
in a maximum matching equals the number of vertices in a minimum vertex cover.
Proof. Let G(V, E) be the bipartite graph with partitions X, Y . Let k be the size
of a maximum matching, M. Then, the size of the minimum vertex cover is at
least k. Since, no two edges in a matching has a common vertex, a vertex cover
should consist of at least one vertex for each edge.
What remains is to show that there exists a vertex cover of k. For this, we rst
transform the given graph as below. Construct the new graph G

(V

, E

), where V

consists of all the vertices from G along with two new vertices s, t corresponding
to source and sink and add edges to E

as below:
(1) u X, v Y add a directed edge (u, v)
(2) u X add the directed edge (s, u)
15
(3) v Y add the directed edge (v, t)
Assign unit capacities to all the edges in E

.
e
f
d a
b
c
(a) Original Graph
s
a
b
d
e
f
t
c
(b) Transformed
Figure 2.2: (a) A bipartite graph with partitions {a, b, c} and {d, e, f}. (b) The resulting
network after transformation.
We now show that the no of pairwise internally-disjoint s, t-paths in G

is
equal to the size of the maximum matching in G. Since no two edges in M have
a common vertex, all the s, t-paths corresponding to the edges in M are pairwise
internally-disjoint. There can not be any other path that is internally-disjoint to
all these paths. Because, if there exists such a path, say s xy t, then the size
of matching M can be increased by adding the edge (x, y). Note that every path
in G

is of length three. This contradicts the fact that M is a maximum matching

in G. Therefore, number of pairwise internally-disjoint paths in G

is k.
By applying Mengers Theorem(Vertex version) to G

, the size of minimum

s, t-separator, say W, is k. We show that W forms a vertex cover in G. Since, for
every edge (x, y) G there exists a s, t-path in G

. For every s, t-path, sxyt,

the s, t-separator consists of either x or y or both. So, W forms a vertex cover of
size k in G.
Chapter 3
Max-ow Min-cut Theorem and Konigs Theorem using Total
Unimodularity
In this chapter, we will use a dierent proof technique to prove the Konigs
theorem and Max-ow Min-cut theorem. We consider the linear programming
formulation of the problem and show that the optimal values of primal and dual
are equal. We use the total unimodularity property of coecient matrix and the
fundamental theorem of duality in linear program to drive this equivalence. The
total unimodularity of the coecient matrix helps in determining the integrality
of the solution. The proof was given by A Chandra Babu et al.[2]. We have lled
a few missing links in this material.
3.1 Preliminaries
Denition 3.1.1. (Unimodular Matrix) A matrix M over real numbers is said to
be unimodular if every square sub matrix of M has determinant equal to 0,1 or -1.
Examples of totally unimodular matrices,
_

_
1 1 0
0 1 1
0 0 1
_

_
_

_
1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1
_

_
17
_

_
1 1 0 0 0 1
1 0 1 1 0 0
0 1 1 0 1 0
0 0 0 1 1 1
_

_
Denition 3.1.2. (Linear Program) Let P be a maximization problem. Consider
this as the primal. Then the linear program formulation can be given as
Maximize c
T
x
Subject to Ax b
x 0
(3.1.1)
Where A is a m n matrix and c, x, b are column vectors of order n, n, m
respectively. c
T
denotes the transpose of c.
A =
_

_
a
11
a
12
. . . a
1n
a
21
a
22
. . . a
2n
.
.
.
.
.
.
.
.
.
.
.
.
a
m1
a
m2
. . . a
mn
_

_
x =
_

_
x
1
x
2
.
.
.
x
n
_

_
c =
_

_
c
1
c
2
.
.
.
c
n
_

_
b =
_

_
b
1
b
2
.
.
.
b
n
_

_
Denition 3.1.3. (Dual of a Linear Program) The dual of the LPP(P), say D(P),
can be given as
Minimize b
T
y
Subject to A
T
y c
y 0
18
Where y is a column vector of order m.
y =
_

_
y
1
y
2
.
.
.
y
m
_

_
Theorem 5. (Weak Duality) For any feasible solutions x

and y

of P and D(P),
c
T
x

is always less than or equal to b

T
y

.
Proof. For a feasible x

and y

of P and D(P), the inequality constraints will be

satised. Consider the inequalities in P and multiply with y
T
on both sides.
Ax

b
y
T
Ax

y
T
b
(3.1.2)
Now, consider the objective function c
T
x

of P. We have,
A
T
y

c
(A
T
y

)
T
x

c
T
x

y
T
Ax

c
T
x

(3.1.3)
Now, from equations 3.1.2 and 3.1.3 we have
c
T
x

y
T
Ax

y
T
b (3.1.4)
19
We can see that y
T
b = b
T
y

. Since,
y
T
b =
_
y
1
y
2
. . . y
m
_

_
b
1
b
2
.
.
.
b
n
_

_
=
_
b
1
y
1
+b
2
y
2
+. . . +b
m
y
m
_
=
_
b
1
b
2
. . . b
n
_

_
y
1
y
2
.
.
.
y
m
_

_
= b
T
y

(3.1.5)
Therefore,
c
T
x

y
T
Ax

b
T
y

(3.1.6)
Theorem 6. (Strong Duality) [?] If the primal P has an optimal solution, x

,
then the dual D(P) also has an optimal solution, y

, such that c
T
x

= b
T
y

.
Theorem 7. (Unimodular LP) For the LPP(P), if A is a unimodular matrix and
b is integral then some optimal solution is integral.
The proof of the above two theorems is beyond the scope of this report. The
proof of the Theorem 6 is found in [?] and the proof of the Theorem 7 is found in
[2].
3.2 Konigs Theorem
Theorem 8. (Konigs theorem,[1931]) In any bipartite graph, the number of edges
in a maximum matching equals the number of vertices in a minimum vertex cover.
20
Proof. Let G(V, E) be a bipartite graph and X, Y be the two partitions of V . Let
|X| = m, |Y | = n and {1, 2, . . . , n} X,{m+ 1, m+ 2, . . . , m+n} Y .
If E is empty, then the size of both the minimum vertex cover and maximum
matching is zero. Without loss of generality we shall assume that E is non empty
and |E| = r. We shall introduce two variables p and q corresponding to every
vertex and edge respectively. The linear program(P) for nding the maximum
matching is as follows,
Maximize

(i,j)E
q
ij
(3.2.1a)
Subject to

j:(i,j)E
q
ij
1 i X (3.2.1b)

i:(i,j)E
q
ij
1 j Y (3.2.1c)
q
ij
{0, 1} (i, j) E (3.2.1d)
The rst two constraints imply that at most one edge can be selected corre-
sponding to every node.
Here, the given problem is an Integer Linear Program. We relax the inte-
grality constraints on q
ij
so that it can take on decimal values. The resulting LPP,
say P

, is
Maximize

(i,j)E
q
ij
Subject to

j:(i,j)E
q
ij
1 i X

i:(i,j)E
q
ij
1 j Y
q
ij
0 (i, j) E
(3.2.2)
Let A be the coecient matrix of the LPP(P

). By Theorem 3.2.1 below,

21
A is unimodular. Hence by Lemma 7, there exists an optimal integral solution to
P

. Obviously, every element of the solution is either 0 or 1. Otherwise it would

violate the constraints. Hence optimal solution of P

is also optimal solution of P.

This is the size of the maximum matching in G.
Consider the LP in equation 3.2.2. We shall derive the dual of this LP
formulation. Let p
i
and p
j
where i X, j Y be the dual variables corresponding
to the rst and second set of constraints, respectively. The matrices corresponding
to the primal formulation 3.1.1 and dual formulation 3.1.3 problems are as follows,
x =
_

_
q
ij
.
.
.
.
.
.
.
.
.
_

_
c =
_

_
1
1
.
.
.
1
_

_
b =
_

_
1
1
.
.
.
1
_

_
y =
_

_
p
1
.
.
.
p
m
p
m+1
.
.
.
p
m+n
_

_
(3.2.3)
Where the order of the matrices are, x is r 1, c is r 1, b is (m + n) 1
and y is (m + n) 1. The optimization function with respect to denition 3.1.3
can be given as,
b
T
y =
_

_
1
1
.
.
.
.
.
.
.
.
.
1
_

_
T

_
p
1
.
.
.
p
m
p
m+1
.
.
.
p
m+n
_

_
(3.2.4)
From Theorem 3.2.1, we know the coecient matrix is unimodular. In the
matrix A, with respect to a variable q
ij
, every column has exactly two ones corre-
sponding to i
th
and j
th
rows. Thus, in A
T
every row has two ones, one in the i
th
22
column and other in the j
th
column. So, the set of constraints, A
T
y c, for the
dual can be given as follows,
p
i
+p
j
1 (i, j) E (3.2.5)
Along with the non-negativity constraints, the dual can be given as,
Minimize

iX
p
i
+

jY
p
j
Subject to,
p
i
+p
j
1 (i, j) E
p
i
0 i X
p
j
0 j Y
(3.2.6)
This actually is the LP formulation for the relaxed minimum vertex cover
problem that we see below.
Now, consider the problem for nding the minimum vertex cover in G. Then
the LPP formulation, say Q, is given as follows,
Minimize

iX
p
i
+

jY
p
j
Subject to,
p
i
+p
j
1 (i, j) E
p
i
{0, 1} i X
p
j
{0, 1} j Y
(3.2.7)
The above given problem is an ILP(Integer Linear Program). We shall relax
the integrality constraints on x
i
, x
j
in order to obtain a LP. Let Q

be the new LP.

23
The new LP is as follows,
Minimize

iX
p
i
+

jY
p
j
Subject to,
p
i
+p
j
1 (i, j) E
p
i
0 i X
p
j
0 j Y
(3.2.8)
Let B be the coecient matrix of Q

. It is easy to see that B is the transpose

of A. Since A is unimodular, B is also unimodular. Therefore, by Theorem 3.2.1,
Q

has an optimal integral solution in which all p

i
and q
j
are either 0 or 1. Hence,
optimal solution to Q

is also optimal solution to Q. That is equal to the size of

the minimum vertex cover.
Now, by the Theorem 6, both the problems have same optimal solution.
Lemma 3.2.1. The coecient matrix A of the LPP (P

) is totally unimodular.
Proof. Clearly, A has m + n rows and r columns in which rst m equations con-
tribute the rst m rows and second n equations contribute the remaining n rows.
Each column of A has exactly two 1s, one in the rst m rows and one in the last
n rows. All other elements are 0. Let D be a square sub-matrix of order k. We
will prove the theorem by using induction on k.
Clearly, for k = 1, |D| = 0 or 1. Assume that all square sub-matrices of
order k 1 have determinant equal to 0, 1 or 1. We shall consider dierent cases,
(1) If D has at least one column containing only zeros, then |D| = 0.
(2) If D has at least one column containing only one one. Then |D| = |E|,
where E is the sub-matrix obtained from D by deleting the corresponding
column and the row containing the one. By induction, |E| = 0, 1 or 1.
Hence |D| = 0, 1 or 1.
24
(3) If every column of D has exactly two 1. In this case, the rst 1 comes from
the rst m rows of A and the second 1 comes from the later n rows of A.
Strictly, every column has exactly a single 1 within the rows that belong
to the rst m rows of A. So, the by performing row addition on all these
rows, we obtain a row with all ones. The same argument applies for the
remaining rows of D that came from the last n rows of A. Hence the rows
of D are linearly dependent and |D| = 0.
Hence A is unimodular.
3.3 The Max-Flow Min-Cut Theorem
Theorem 9. (Ford-Fulkerson, 1956) In a Network G, let f be any maximum ow
in G, then a cut (A, B) for which f(A, B) = c(A, B)
Proof. Let N(V, E, c, s, t) be the network with |V | = n and |E| = m. We shall
write the linear programming formulation for the maximum ow.
We want to nd the maximal ow that can be sent from the source vertex s
to the sink vertex t. Let v be the value of any ow from s to t and x
ij
be the ow
sent along the arc (i, j). Let the vertices be labeled using integers 1 to n such that
the source s is labeled as 1 and the sink t is labeled with n. Then the LPP(R)
25
corresponding to the maximum ow is,
Maximize v
Subject to

(i,j)E
x
ij

(k,i)E
x
ki
v = 0 if i = 1

(i,j)E
x
ij

(k,i)E
x
ki
+v = 0 if i = n

(i,j)E
x
ij

(k,i)E
x
ki
= 0 if i = 2, 3, . . . , n 1
x
ij
c
ij
(i, j) E
x
ij
0 (i, j) E
v 0
(3.3.1)
The rst constraint imply that the net ow out of the source vertex 1 is equal
to v and the second constraint imply that the net ow into the sink vertex n is
v. The third constraint imply that the total ow into any intermediate vertex is
equal to the total outow of that vertex. By relaxing the equality in these three
constraints, we will obtain the following inequalities.

(i,j)E
x
ij

(k,i)E
x
ki
v 0 if i = 1

(i,j)E
x
ij

(k,i)E
x
ki
+v 0 if i = n

(i,j)E
x
ij

(k,i)E
x
ki
0 if i = 2, 3, . . . , n 1
(3.3.2)
Let R

be the resulting LPP. In any optimal solution of R

, the above three

inequalities should satisfy with equality. Otherwise, by adding all the LHS and
RHS we get 0 < 0, a contradiction. Since, on the LHS for every arc (i, j), x
ij
is
added once and subtracted once, so the sum will result in a zero. So, they will
satisfy with equality. Therefore, the optimal solution of R

will also be the optimal

solution of R.
26
Consider the LP for Max-ow in equation 3.3.1. If we try to convert this
into matrix form, the corresponding matrices will be,
x =
_

_
x
ij
.
.
.
.
.
.
_

_
b =
_

_
0
.
.
.
0
c
ij
.
.
.
_

_
c =
_

_
1
0
0
.
.
.
_

_
(3.3.3)
Now, we shall derive the dual for the above LP in equation 3.3.1. Let u
i
,
where 1 i n, be the dual variables corresponding to the rst three set of
equations (ow constraints) and y
ij
, where (i, j) E, be the dual variables corre-
sponding to the fourth set of constraints (capacity constraints). Then the matrix
y is,
y =
_

_
u
1
.
.
.
u
n
y
ij
.
.
.
_

_
(3.3.4)
Therefore, the objective funtion for the dual can be given as,
b
T
y =
_

_
0
.
.
.
0
c
ij
.
.
.
_

_
T

_
u
1
.
.
.
u
n
y
ij
.
.
.
_

_
(3.3.5)
Consider the coecient matrix, say A, of the LP formulation for Max-ow.
We shall see the properties of this matrix. There will be (n+m) rows corresponding
27
to the (n + m) constraints and m + 1 columns corresponding to m x
ij
variables.
The coecient matrix A, will look as below,
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
1 2 3 . . . m+ 1
1 1 a
2,1
a
3,1
. . . a
m+1,1
2 0 a
2,2
a
3,2
. . . a
m+1,2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
n 1 a
2,n
a
3,n
. . . a
m+1,n
x
ij
0 1 0 . . . 0
.
.
. 0 0 1 . . . 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 0 0 0 . . . 1
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
where a
i,j
= 1 or 0 or 1 2 i (m + 1), 1 j n. This matrix is
actually the transpose of the coecient matrix for the Min-cut LP that we see
below in equation 3.3.6. The objective function of the Min-cut is also the same
as the function in equation 3.3.5. From this, it is easy to see that the dual of the
Max-ow problem is Min-cut.
Now we shall formulate the linear program(T) for nding the minimum cut
capacity as follows,
Minimize

(i,j)E
c
ij
y
ij
Subject to
u
1
+u
n
1
u
i
u
j
+y
ij
0 (i, j) E
u
i
{0, 1} i
y
ij
{0, 1} i, j
(3.3.6)
The solution to the above LPP will result in a cut such that, corresponding
28
to a cut (S,

S) of the network N,
u
i
= 0 if vertex i S
= 1 if vertex i

S
y
ij
= 1 if i S, j

S
= 0 otherwise
(3.3.7)
Now, relax the integrality constraints on u
i
and y
ij
. The resulting LPP(T

)
will be
Minimize

(i,j)E
c
ij
y
ij
Subject to
u
1
+u
n
1
u
i
u
j
+y
ij
0 (i, j) E
u
i
0 i
y
ij
0 i, j
(3.3.8)
In any optimal solution to T, (u
i
, u
j
) = (0.0) or (1, 0) or (1, 1) will imply
y
ij
= 0 and (u
i
, u
j
) = (0, 1) imply y
ij
= 1 and hence T will give the capacity of the
cut (S,

S).
Now, consider the LPP(T

). Clearly, this is the dual of the LPP(R

). Let B
be the coecient matrix of T

. From Lemma 3.3.1 below, B is unimodular. The

column vector b of T

consists either 0s or 1s. Then the optimal solution of T

is
also the optimal solution of T.
Now, by Theorem 6, the optimal values of R

and T

are equal. Hence, the

optimal values of R and T are also equal. Therefore, the maximum ow in the
network is equal to the minimum cut in the network.
Lemma 3.3.1. The coecient matrix B of the LPP(T) is unimodular.
Proof. Consider the matrix B. Clearly B has m+1 rows and n+m columns. Now,
let D and E be two partitions of B, such that D consists of the rst n columns
29
and E consists of the second m columns. The matrix D is of order (m + 1) n
and E is of order (m+ 1) m. The matrix D will be as below,
_
_
_
_
_
_
_
_
_
u
1
u
2
. . . u
n
1 1 0 . . . 1
2 a
2,1
a
2,2
. . . a
2,n
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
m+ 1 a
m+1,1
a
m+1,2
. . . a
m+1,n
_
_
_
_
_
_
_
_
_
where a
i,j
= 1 or 0 or 1 2 i (m + 1), 1 j n. And the matrix E
will be as below,
_
_
_
_
_
_
_
_
_
_
_
_
_
y
ij
. . . . . .
1 0 0 . . . 0
2 1 0 . . . 0
3 0 1 . . . 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
m+ 1 0 0 . . . 1
_
_
_
_
_
_
_
_
_
_
_
_
_
Every row of C contains exactly one 1 and one -1. Every column of D will
contain exactly one 1. Now, we shall prove this by induction on the size of the
sub-matrix. let U be the square sub-matrix of order k of B. For, k=1, the element
can only be either 1 or 0 or -1. Assume that all square sub-matrices of order k 1
have determinant equal to 0, 1 or 1. Considering for k, the dierent cases can be
as,
(1) U consists of at least one column from D. Every column of D has exactly
one 1. Then by deleting that column and the corresponding row, we can
get a matrix of order k 1. Hence, |U| = 1 or 0 or 1.
(2) U consists of columns only from C. Now, if there exists a row with all zeros
or one 1 or one -1, then by induction we can see that the determinant of U
30
will be 1 or 0 or -1. Otherwise, every row of U should contain exactly one
1 and one -1. Then by performing the column addition on all the columns
will result in a column with all zeros. Then |U| = 0.
Chapter 4
The Concurrent Multi-commodity Flow Problem
This chapter consists of introduction to Multi-commodity ow problem. The
Multi-commodity ow problem is a more generalized network ow problem. In
a multi-commodity ow problem there are k 1 commodities each having its
own source and sink pair. Because of the multiplicity of the commodities, the
problem to be optimized can be dened in several ways. Hence, there exists several
variants of the Multi-commodity ow. The problem we are going to discuss is called
Concurrent Multi-commodity Flow Problem(CMFP) [15][12]. In this problem,
every i
th
commodity is assigned a demand D
i
. Our objective is to maximize a
fraction , such that there exists a ow of D
i
units for every commodity i. The
previously studied maximum ow problem is a special case of Multi-commodity
ow problem in which the number of commodities is one(k = 1). Below, we will
give formal denitions and linear program formulation of the problem.
4.1 Preliminaries
Denition 4.1.1. (Multi-commodity Network) A Multi-commodity Network is a
directed graph G(V, E, c) with vertex set V and an arc set E in which every directed
edge (i, j) E, has a non negative capacity c(i, j) 0, c : V V R
+
. There
are k 1 commodities K
1
, K
2
, . . . , K
k
. For each commodity i, there is an ordered
pair (s
i
, t
i
) representing the source and sink of that commodity where (s
i
, t
i
) V V
and s
i
= t
i
.
32
The ow of a commodity is similar to that of a ow in a single-commodity
ow network. Let f
i
represent the ow of the i
th
commodity.
Denition 4.1.2. (Flow of a Commodity) A ow of a commodity is a mapping
f
i
: E R denoted by f
i
uv
or f
i
(u, v), subject to the following constraints:
(1) f
i
(u, v) c(u, v) for each (u, v) E (capacity constraint)
(2) f
i
(u, v) = f
i
(v, u) (skew symmetry)
(3)

vV
f
i
(u, v) = 0 u V \ {s
i
, t
i
} (conservation of ow).
The value of the ow of a commodity i is given by |f
i
| =

wV
f
i
(s
i
, w).
Denition 4.1.3. (Concurrent Multi-commodity Flow Problem) Given a Multi-
commodity Network along with demands D
1
, D
2
, . . . , D
k
corresponding to the k
commodities, the objective is to assign ow to commodities so as to maximize a
fraction such that for every commodity i, the value of the ow of the commodity
|f
i
| is at least D
i
. The assignment should satisfy the following constraints along
with the ow constraints,
k

i=1
f
i
(u, v) c(u, v) (u, v) E (4.1.1)
4.2 Linear Programming Formulation
Below, we give the LP formulation for the Concurrent Multi-commodity Flow
problem(CMFP).
33
Maximize
Subject to
k

i=1
f
i
(u, v) c(u, v) (u, v) E

wV
f
i
(u, w) = 0 1 i k, u V {s
i
, t
i
}

wV
f
i
(s
i
, w) D
i
1 i k
(4.2.1)
The above formulation is very intuitive and straight forward. We will now
see another formulation of the same problem which uses paths. Let P represent the
set of all non-trivial paths in the network and P
j
be the set of paths corresponding
to the commodity j (paths from s
i
to t
i
). Let x() be a variable corresponding to
every path P. Then, the LPP formulation (primal) can be given as,
Maximize
Subject to

:e
x() c(e) e E
D
j

P
j
x() 0 1 j k
x() 0 P
(4.2.2)
The dual problem for the above linear program can be interpreted as assigning
weights(z
j
) to the commodities and lengths(y(e)) to edges such that for any com-
modity i, the length of every path from s
i
to t
i
should be at least z
i
. The length
of a path is given as the sum of the lengths of all the edges in that path. The LPP
34
formulation for the dual is as follows,
Minimize

eE
c(e)y(e)
Subject to

e
y(e) z
j
P
j
, j

1jk
D
j
z
j
1
l(e) 0 e E
z
j
0 j
(4.2.3)
We shall consider the example in the gure 4.1 with two commodities, say K
1
and
K
2
. Let the demands be D
1
and D
2
respectively. First, the primal formulation of
the problem using paths is given. We shall then derive the dual formulation of the
problem.
s
1
s
2
t
1
t
2
e
1
e
2
e
3 e
4
e
5
e
6
e
7
e
8
a
b
c
d
Figure 4.1: An example of a two commodity ow network with unit demands on the
commodities.
In the example gure, there are two paths from source to sink for each of
the commodities K
1
and K
2
. The edges are labeled e
1
, e
2
, . . . , e
8
. Let us name the
paths with respect to edges as follows,
35

1
e
1
e
2
e
3
e
4

2
e
1
e
6
e
7
e
4

3
e
5
e
2
e
3
e
8

4
e
5
e
6
e
7
e
8
(4.2.4)
The paths
1
,
2
belong to commodity K
1
and the paths
3
,
4
belong to
commodity K
2
. The LP formulation for the example is as follows,
Maximize
Subject to
x(
1
) +x(
2
) c(e
1
)
x(
1
) +x(
4
) c(e
2
)
x(
1
) +x(
4
) c(e
3
)
x(
1
) +x(
2
) c(e
4
)
x(
3
) +x(
4
) c(e
5
)
x(
3
) +x(
2
) c(e
6
)
x(
3
) +x(
2
) c(e
7
)
x(
3
) +x(
4
) c(e
8
)
D
1
x(
1
) x(
2
) 0
D
2
x(
3
) x(
4
) 0
x(
i
) 0 1 i 4
(4.2.5)
Now, multiply the constraints with the dual variables(y(e) and z
j
) on both sides.
36
[ x(
1
) +x(
2
) c(e
1
) ] y(e
1
)
[ x(
1
) +x(
4
) c(e
2
) ] y(e
2
)
[ x(
1
) +x(
4
) c(e
3
) ] y(e
3
)
[ x(
1
) +x(
2
) c(e
4
) ] y(e
4
)
[ x(
3
) +x(
4
) c(e
5
) ] y(e
5
)
[ x(
3
) +x(
2
) c(e
6
) ] y(e
6
)
[ x(
3
) +x(
2
) c(e
7
) ] y(e
7
)
[ x(
3
) +x(
4
) c(e
8
) ] y(e
8
)
[ D
1
x(
1
) x(
2
) 0 ] z
1
[ D
2
x(
3
) x(
4
) 0 ] z
2
(4.2.6)
By combining the equations on the left hand and the right hand sides we get the
following inequality,
[ x(
1
) +x(
2
)] y(e
1
) + [ x(
1
) +x(
4
)] y(e
2
)+
[ x(
1
) +x(
4
)] y(e
3
) + [ x(
1
) +x(
2
)] y(e
4
)+
[ x(
3
) +x(
4
)] y(e
5
) + [ x(
3
) +x(
2
)] y(e
6
)+
[ x(
3
) +x(
2
)] y(e
7
) + [ x(
3
) +x(
4
)] y(e
8
)+
[ D
1
x(
1
) x(
2
) ] z
1
+
[ D
2
x(
3
) x(
4
) ] z
2

8

i=1
c(e
i
)y(e
i
)
(4.2.7)
37
Now, represent the inequality in terms of x(),
[ y(e
1
) +y(e
2
) +y(e
3
) +y(e
4
) z
1
] x(
1
)+
[ y(e
1
) +y(e
6
) +y(e
7
) +y(e
4
) z
1
] x(
2
)+
[ y(e
5
) +y(e
6
) +y(e
7
) +y(e
8
) z
2
] x(
3
)+
[ y(e
5
) +y(e
2
) +y(e
3
) +y(e
8
) z
2
] x(
4
)+
[ D
1
z
1
+D
2
z
2
]
8

i=1
c(e
i
)y(e
i
)
(4.2.8)
Now, with respect to the objective function of the primal, the dual can be formu-
lated as below,
Minimize
8

i=1
c(e
i
)y(e
i
)
Subject to
y(e
1
) +y(e
2
) +y(e
3
) +y(e
4
) z
1
0
y(e
1
) +y(e
6
) +y(e
7
) +y(e
4
) z
1
0
y(e
5
) +y(e
6
) +y(e
7
) +y(e
8
) z
2
0
y(e
5
) +y(e
2
) +y(e
3
) +y(e
8
) z
2
0
D
1
z
1
+D
2
z
2
1
y(e
i
) 0 1 i 8
z
j
0 1 j k
(4.2.9)
This is the resulting LPP formulation for the dual of the example we consid-
ered in gure 4.1. The dual we have given in equation 4.2.3 is a generalization of
this resulting formulation.
The Multi commodity ow problem is very well studied in combinatorics.
Unlike single commodity ow, the structural properties of this problem are not
well known when the number of commodities is greater than two(k > 2). This
38
problem can be solved in polynomial time using linear programming. However,
the problem of nding an integer ow is NP-Complete when k 2.
Chapter 5
Conclusion
In this thesis, we reviewed the classical Max-ow Min-cut theorem and its
proof using Ford-Fulkerson algorithm. We have also presented the Konigs the-
orem and Mengers theorem as consequences of the Max-ow Min-cut theorem.
While the above proofs are very well established, they are proved in an algorithmic
perspective. In the third chapter, we have presented the proofs for Konigs theo-
rem and Ma-ow Min-cut theorem using a complete dierent technique based on
the total unimodularity property of the coecient matrix in their linear program
formulation. Finally, we have briey discussed about Multi-commodity ow and
the Concurrent Multi-commodity Flow Problem(CMFP).
Although these results and the formulations are not new, an attempt was
made to present complete proofs for the results from rst principles, and the mate-
rial does not seem to be consolidated and presented elsewhere in an easily accessible
form.
Many more primal-dual relations exist in graph theory and the approach
generalize to investigation into these relations and discovering LP based proofs for
those Min-Max relations in graph theory. Hence, this approach is a general tool
and the results presented here are just sample cases.
This study gives an insight into dierent techniques and tools that can be
used to prove primal-dual relations in Graph theory. Though we have not estab-
lished any new results, the treatment given in this thesis gives us a good scope of
40
applying these techniques in order to establish new results in both graph theory
and combinatorics. We have also attempted to apply a technique called Lagrangian
relaxation [7] from linear programming to some of these relations in order to gain
some insights into its eectiveness. But, the work related that is not included in
this thesis as it did not yield insightful results. A possibility is to apply dierent
techniques in combination and try to investigate the outcome which could lead to
interesting observations.
Bibliography
[1] Ron Aharoni and Eli Berger. Mengers theorem for innite graphs. Inventiones
Mathematicae, 176(1):162, 2009.
[2] A Chandra Babu, P.V. Ramakrishnan, and C.R. Seshan. New proofs of konig-
egervary theorem. and maximal ow-minimal cut capacity. theorem using o.r.
techniques., August 1990.
[3] Thomas Cormen. Introduction to algorithms. The MIT Press, Cambridge
Mass., 2. ed. edition, 2001.
[4] Thomas Cormen. Introduction to algorithms. The MIT Press, Cambridge
Mass., 2. ed. edition, 2001.
[5] L. R. Ford and D. R. Fulkerson. Maximal ow through a network. Canadian
Journal of Mathematics, 8:399404, January 1956.
[6] Anders Forsgren. Optimization online - an elementary proof of optimality
conditions for linear programming. Technical Report TRITA-MAT-2008-OS6,
Department of Mathematics, Royal Institute of Technology (KTH), SE-100
44 Stockholm, Sweden, June 2008.
[7] Arthur M. Georion. Lagrangian relaxation for integer programming. In
Michael J unger, Thomas M. Liebling, Denis Naddef, George L. Nemhauser,
William R. Pulleyblank, Gerhard Reinelt, Giovanni Rinaldi, and Laurence A.
Wolsey, editors, 50 Years of Integer Programming 1958-2008, pages 243281.
Springer Berlin Heidelberg, Berlin, Heidelberg, 2010.
[8] Alon Itai. Two-Commodity ow. Journal of the ACM, 25(4):596611, October
1978.
[9] Denes K onig. Gr afok es alkalmaz asuk a determin ansok es a halmazok
elmeletere. Matematikai es Termeszettudom anyi

Ertesto, 34:104119, 1916.
[10] Eugene Lawler and Eugene Lawler. Combinatorial optimization: Networks
and matroids. In 4.5. Combinatorial Implications of Max-Flow Min-Cut
Theorem, 4.6. Linear Programming Interpretation of Max-Flow Min-Cut
Theorem, pages 117120. Dover, 2001.
42
[11] Eug`ene L. Lawler. Combinatorial Optimization: Networks and Matroids.
Courier Dover Publications, March 2001.
[12] Tom Leighton and Satish Rao. Multicommodity max-ow min-cut theorems
and their use in designing approximation algorithms. J. ACM, 46(6):787832,
November 1999.
[13] L aszl o Lov asz and M. D Plummer. Matching theory. North-Holland : Elsevier
Science Publishers B.V. ; Sole distributors for the U.S.A. and Canada, Elsevier
Science Pub. Co., Amsterdam; New York; New York, N.Y., 1986.
[14] Christos Papadimitriou. Combinatorial optimization : algorithms and
complexity. Dover Publications, Mineola N.Y., 1998.
[15] Farhad Shahrokhi and D. W. Matula. The maximum concurrent ow problem.
J. ACM, 37(2):318334, April 1990.
[16] Douglas B. West. Introduction to Graph Theory. Prentice Hall, 2 edition,
September 2000.