ISE 632: Network Flows and Combinatorial Optimization

Lecture 3: Total unimodularity and applications

Prof. J. G. Carlsson

Epstein Department of Industrial and Systems Engineering, University of Southern

California
Suggested reading

I B Korte and J Vygen. Combinatorial optimization.

Springer, 2002
Total unimodularity

I We just saw that the linear relaxation of the maximum-weight

matching problem has a {0, 1} solution because all corners of the
feasible region have integer solutions
I We say that the integrality gap (the difference between the LP
and the IP) is zero
I This is a special case of a general property called total
unimodularity :

Definition
A matrix A is said to be totally unimodular (TU) if each
sub-determinant of A is either 0, 1, or −1
I It turns out that an integer matrix A is totally unimodular if and
only if the set {x : Ax ≤ b , x ≥ 0} has integer corners for all
integer vectors b
Proof ( =⇒ )

I Suppose that A ∈ Rm×n is TU and let v be a corner of

{x : Ax ≤ b}
0 0
I Since v is a corner, we can write A v = b for a square submatrix
0 0
A ∈ Rn×n and some entries b
I Cramer’s rule says that
0
det(Ai )
vi =
det(A0 )
0 0 0
where Ai is the matrix where we replace the ith column of A by b
I The numerator is a product of integers and the denominator is ±1
I The other direction is less interesting/useful; see [1]

[1] B Korte and J Vygen. Combinatorial optimization.

Springer, 2002.
Facts about determinants
I det A = det AT
I If A and B have the same (square) size, then
det(AB) = det A det B
 
I det a b = ad − bc
c d
I Exchanging rows (or columns) reverses the sign:
   
a1 a1
 ..   .. 

 .  

 . 

 ai    aj 
..  = − det  .. 
  
det 

 . 

 .  

 aj  

 ai  
 ..   .. 
 .   . 
an an
I A row/column of all zeros implies det A = 0
I A row/column with all zeros except for a single 1 does not change
things:  
1 0
det = det M
a M
Facts about determinants

 
I If A ∈ Rm1 ×n and B ∈ Rm2 ×n are TU, is A
necessarily TU?
B
Facts about determinants

 
I If A ∈ Rm1 ×n and B ∈ Rm2 ×n are TU, is A
necessarily TU?
B



I No way! Use e.g. A = 1 −1 and B = 1 1
   
I The set x : A 2
x≤ has (3/2, −1/2) as a corner
B 1
Corollaries

I A is TU ⇐⇒ AT is TU
I A is TU ⇐⇒ (A, I ) is TU
I A is TU ⇐⇒ the corners of {x : Ax ≤ b, x ≥ 0} are always
integers if b is an integer
I A is TU =⇒ the corners of {x : Ax = b, x ≥ 0} are always
integers if b is an integer
I The dual problem of

maximize cT x s.t. Ax ≤ b, x ≥ 0 ,
x

which is
minimize bT y s.t. AT y ≥ c ,
y

is also integral, if b and c are integral

Equivalent condition

An equivalent expression of total unimodularity is:

Theorem
(Ghouila-Houri 1962) An m × n matrix A is totally unimodular if and
only if, for every R ⊆ {1, . . . , m}, there is a partition R = R1 ∪ R2 such
that X X
aij − aij ∈ {−1, 0, 1}
i∈R1 i∈R2

for all j ∈ {1, . . . , n}.

Example The “consecutive ones” property within each column
Consecutive ones example

I We have time periods 1, . . . , m and we need to schedule workers

for these time periods
I Each period i needs di workers
I Workers should only work on consecutive periods because it
would be a pain to leave and come back
I Each shift j looks like

aj = (0, . . . , 0, 1, . . . , 1, 0, . . . , 0)

and there could be many different shifts (say N in total), which we

pay cj for
I Let A be the matrix whose columns are all the entries aj
Consecutive ones example

I The LP is

minimize cT x s.t.
x
Ax ≥ d
x ≥ 0

where xj indicates how many workers are assigned to shift j

I This has the consecutive ones property, thus TU, and the LP is
integral
Example

I The bipartite matching polytope is totally unimodular:

n
X
xij = 1
i=1
Xn
xij = 1
j=1
xij ≥ 0

I This says that the integer program for a bipartite matching

problem is solvable as an LP (we already proved this using
Birkhoff-von Neumann)
Matching polytope

 
1 1 1
 .. .. .. 

 . . ··· . 


 1 1 1 

 1
A= ··· 1 


 1 ··· 1 

 .. 
 . 
1 ··· 1
Example

I The incidence matrix of any directed network is TU

I Recall that the incidence matrix for a directed network with n
vertices v1 , . . . , vn and k edges e1 , . . . , ek is an n × k matrix B
such that bij = 1 if vertex i is the origin of edge j, bij = −1 if
vertex i is the destination of edge j, and bij = 0 otherwise
Example

I The incidence matrix of any directed network is TU

I Recall that the incidence matrix for a directed network with n
vertices v1 , . . . , vn and k edges e1 , . . . , ek is an n × k matrix B
such that bij = 1 if vertex i is the origin of edge j, bij = −1 if
vertex i is the destination of edge j, and bij = 0 otherwise
I Just set R1 = R and R2 = ∅!
Maximum flow

I Given a directed, weighted graph G = (V , E , W ) and a pair of

vertices s and t
I Think of the edge weights wij as the “capacity” of that edge
I What’s the largest amount of “flow” we can send from s to t,
subject to the capacity constraints?
3
3
a b
4
6

s 1 2 7 t
2
6
2 5
c d
2
Maximum flow

I Let xij denote the amount of flow that we send across edge (i, j);
our problem is
X
maximize xsi s.t.
(s,i)∈E
xij ≤ wij
X X
xij = xji ∀i ∈
/ {s, t}
(i,j)∈E (j,i)∈E

xij ≥ 0 ∀ (i, j) ∈ E

I The above polyhedron is TU because it is described by the

incidence matrix of A
I More generally, the set of all single-commodity flows has a TU
matrix
Minimum-cost flow

I Let xij denote the amount of flow that we send across edge (i, j);
our problem is
X
minimize cij xij s.t.
(i,j)∈E
X
xsj = q
j
xij ≤ wij
X X
xij = xji ∀i ∈
/ {s, t}
(i,j)∈E (j,i)∈E

xij ≥ 0 ∀ (i, j) ∈ E

I The above polyhedron is TU because it is described by the

incidence matrix of A
Maximum flow with extra edge

I Let’s add in an extra edge with capacity ∞ from t to s:

3 3
a b a b
4 4
6 6

s 1 2 7 t s 1 2 7 t

6 6
5 5
c d c d
2 2

I Now we just need to maximize the amount of stuff from t to s,

and impose balance constraints everywhere (this makes the
formulation easier)
Maximum flow

maximize xts s.t.

X
X X
xij − xji =0 ∀i ∈ V
j:(i,j)∈E j:(j,i)∈E

xij ≤ wij (i, j) ∈ E

xij ≥ 0 (i, j) ∈ E

Just for fun (to make the dual easier), let’s remove the constraint that
xts ≥ 0 since we’re maximizing xts anyway (so it won’t affect anything)
Maximum flow in matrix form

maximize cT x s.t.
X
Ax = 0
x ≤ w
x ≥ 0 (except xts , for fun)

here A is the incidence matrix of the network and c is all zeros except
for the position corresponding to xts , which has a 1 in it
I The above constraint matrices are TU, thus we have an integral
solution if w is integral
Reminder: primal-dual relationship

Primal minimize cT x maximize bT y Dual

A AT
≥ bi ≥0
constraints ≤ bi ≤0 variables
= bi free
≥0 ≤ cj
variables ≤0 ≥ cj constraints
free = cj
Dual in matrix form

minimize wT z s.t.
y,z
 

y
AT I ≥ c
z
z ≥ 0

the TU property is preserved

Dual of maximum flow

X
minimize wij zij s.t.
y,z
(i,j)∈E

yi − yj + zij ≥ 0 (i, j) ∈ E \ (t, s)

yt − ys + zts = 1
zij ≥ 0 (i, j) ∈ E
Dual of maximum flow

X
minimize wij zij s.t.
y,z
(i,j)∈E

zij ≥ yj − yi (i, j) ∈ E \ (t, s)

zts = 1 − yt + ys
zij ≥ 0 (i, j) ∈ E

I We can assume WLOG that ys = 0

I Remember that wts = ∞, therefore we should have zts = 0,

whence
0 = 1 − yt + ys = 1 − yt =⇒ yt = 1
Dual of maximum flow

X
minimize wij zij s.t.
y,z
(i,j)∈E

zij ≥ yj − yi (i, j) ∈ E \ (t, s)

zts = 1 − yt + ys
zij ≥ 0 (i, j) ∈ E

I We can assume WLOG that ys = 0

I This is because we can set y 7→ y + c for any constant c and have
the same objective value
I Remember that wts = ∞, therefore we should have zts = 0,
whence
0 = 1 − yt + ys = 1 − yt =⇒ yt = 1
Dual of maximum flow

X
minimize wij zij s.t.
y,z
(i,j)∈E

zij ≥ yj − yi (i, j) ∈ E \ (t, s)

ys = 0
yt = 1
zij ≥ 0 (i, j) ∈ E

I We’d like to have yi ≥ yj for all the edges (i, j) ∈ E , but since
ys = 0 and there are paths from s to t, this can’t always be the
case
Dual of maximum flow

X
minimize wij max{yj − yi , 0} s.t.
y
(i,j)∈E
ys = 0
yt = 1

I We’d like to have yi ≥ yj for all the edges (i, j) ∈ E , but since
ys = 0 and there are paths from s to t, this can’t always be the
case
I Claim that the optimal solution has 0 ≤ yi ≤ 1; why?
Dual of maximum flow

X
minimize wij max{yj − yi , 0} s.t.
y
(i,j)∈E
ys = 0
yt = 1

I We’d like to have yi ≥ yj for all the edges (i, j) ∈ E , but since
ys = 0 and there are paths from s to t, this can’t always be the
case
I Claim that the optimal solution has 0 ≤ yi ≤ 1; why?
I Take any nodes where yi > 1 and set them equal to 1; all this
does is decrease the differences yj − yi (same argument for yi < 0)
Dual of maximum flow

X
minimize wij max{yj − yi , 0} s.t.
y
(i,j)∈E
ys = 0
yt = 1

I We’d like to have yi ≥ yj for all the edges (i, j) ∈ E , but since
ys = 0 and there are paths from s to t, this can’t always be the
case
I Claim that the optimal solution has 0 ≤ yi ≤ 1; why?
I Take any nodes where yi > 1 and set them equal to 1; all this
does is decrease the differences yj − yi (same argument for yi < 0)
I By TU property, there exists optimal y∗ such that yi∗ ∈ {0, 1}
Minimum cut problem

X
minimize wij max{yj − yi , 0} s.t.
y
(i,j)∈E
ys = 0
yt = 1
yi ∈ {0, 1} ∀i ∈ V

I The above problem (which we can solve as an LP) is called the

minimum cut problem
Max-flow min-cut theorem

3
3 3
a b a b
4 4
6 6

s 1 2 7 t s 1 2 7 t
2
6 6
5 2 5
c d c d
2 2

I We want a subset C of vertices that contains s (and doesn’t

contain t) that minimizes the total weight of all the edges that go
from C to V \ C
I The preceding result says that the smallest possible “s − t cut”
has the same objective value as the maximum flow (both are 7 in
the diagram above)
Maximum matching and maximum flow

a v

b w

c x

d y

e z

I To find a maximum bipartite matching, just add s and t to a

bipartite network, orient the edges, and insert new edges from s to
U and W to t with capacity 1
I Maximum flow ⇐⇒ maximum matching
I We could use weights of 1 for the middle edges instead of ∞
Maximum matching and maximum flow

a ∞ v
∞ ∞
1 w
b 1
1 ∞
s 1 c ∞ x 1 t
1 ∞ 1
∞
1 d ∞ y 1
∞
e ∞ z

I To find a maximum bipartite matching, just add s and t to a

bipartite network, orient the edges, and insert new edges from s to
U and W to t with capacity 1
I Maximum flow ⇐⇒ maximum matching
I We could use weights of 1 for the middle edges instead of ∞
Easy proof of Hall’s theorem
a v

b w

c x

d y

e z

I Obviously, if there exists S with |N(S)| < |S|, then no perfect

matching exists (same as before)
I On the other hand, suppose a maximum matching (equivalently, a
maximum flow) has value < n
I Then there is a cut C with value < n as well; we claim that
setting S = C ∩ U is a constricting set
I Say there are k blue nodes in the cut, so (n − k) edges from s to
U are counted
I We have (n − k) + |N(S)| < n =⇒ |N(S)| < k, as desired
Easy proof of Hall’s theorem
a ∞ v
∞ ∞
1 w
b 1
1 ∞
s 1 c ∞ x 1 t
1 ∞ 1
∞
1 d ∞ y 1
∞
e ∞ z

I Obviously, if there exists S with |N(S)| < |S|, then no perfect

matching exists (same as before)
I On the other hand, suppose a maximum matching (equivalently, a
maximum flow) has value < n
I Then there is a cut C with value < n as well; we claim that
setting S = C ∩ U is a constricting set
I Say there are k blue nodes in the cut, so (n − k) edges from s to
U are counted
I We have (n − k) + |N(S)| < n =⇒ |N(S)| < k, as desired
Easy proof of Hall’s theorem
a ∞ v
∞ ∞
1 w
b 1
1 ∞
s 1 c ∞ x 1 t
1 ∞ 1
∞
1 d ∞ y 1
∞
e ∞ z

I Obviously, if there exists S with |N(S)| < |S|, then no perfect

matching exists (same as before)
I On the other hand, suppose a maximum matching (equivalently, a
maximum flow) has value < n
I Then there is a cut C with value < n as well; we claim that
setting S = C ∩ U is a constricting set
I Say there are k blue nodes in the cut, so (n − k) edges from s to
U are counted
I We have (n − k) + |N(S)| < n =⇒ |N(S)| < k, as desired
Menger’s theorem

Theorem
In any directed network with nodes s and t, the maximum number of
edge-disjoint paths from s to t is equal to the minimum number of
edges whose removal separates s from t.
Menger’s theorem

Theorem
In any directed network with nodes s and t, the maximum number of
edge-disjoint paths from s to t is equal to the minimum number of
edges whose removal separates s from t.

Proof.
This is a special case of max-flow min-cut where all edges have weight
1.
Total dual integrality (TDI)

I Total unimodularity is a property of a matrix, not a linear system

I A weaker condition (but still a very strong one!) is total dual
integrality (TDI): consider the primal-dual pairs
maximize c T x s.t. minimize b T y s.t.
(1) (2)
Ax ≤ b AT y = c
y ≥0

Definition
The system Ax ≤ b is totally dual integral (TDI) if for all integer c,
the dual problem (2) has an integral solution y ∗ whenever the optimal
value is finite.

Theorem
If Ax ≤ b is TDI and b is integral, then all the corners of the polytope
{x : Ax ≤ b} are integral.
References I

[1] B Korte and J Vygen.

Combinatorial optimization.
Springer, 2002.