Prof.

Friedrich Eisenbrand Due Date: March 09, 2009

Martin Niemeier Discussions: March 10, 2009

Introduction to Discrete Optimization

Spring 2009
Solutions 3

Exercise 1
Consider a school district with I neighborhoods, J schools and G grades at each school. Each
school j has a capacity of C j g for grade g . In each neighborhood i , the student population of
grade g is S i g . Finally the distance of school j from neighborhood i is d i j . Formulate a linear
programming problem whose objective is to assign all students to schools, while minimizing
the total distance traveled by all students. (You may ignore the fact that numbers of students
must be integer.)
Solution
Let x i j g be the amount of students from neighborhood i of grade g travelling to school j .
Then for an assignment of students to schools, the total distance travelled by all students is
given as XX X
di j xi j g .
i I j J g G

For a feasible assignment, every student of every neighborhood and grade must be as-
signed to a school, this gives the constraint
X
x i j g = S i g i I , g G.
j J

The number of students each school can take of the respective grades is bounded by C j g ,
thus X
xi j g C j g j J , g G
i I

must hold. Finally there can be no negative numbers of assignments: x 0. This gives the
following linear program:
P P P
min i I j J g G d i j x i j g
P
subject to j J xi j g = S i g i I , g G
P
i I xi j g C j g j J,g G
x 0

Exercise 2
Consider the vectors

3 1 2 2 1
x1 = 1 , x2 = 2 , x3 = 0 , x4 = 4 , x5 = 1 .

2 5 1 3 1

1
 Let A = {x 1 , . . . , x 5 }. Find two disjoint subsets A 1 , A 2 A such that

conv(A 1 ) conv(A 2 ) 6= ;.

Hint: Recall the proof of Radons lemma

Solution
Since we have 5 points in R3 , Radons lemma states that the subsets A 1 and A 2 exist. To
compute them, we review the proof of Radons lemma.
We construct the set

0 x1 x2 x3 x4 x5
A = , , , ,
1 1 1 1 1

The vectors of A 0 are linearly dependent and we can compute a nontrivial linear combination
of the all zero vector.
5
xi
i
X
0= .
i =1
1
As shown in the proof, if the define sets P := {i : i 0} and N := {i : i < 0}, then the sets
A 1 := {x i : i P } and A 2 := {x i : i N } have the desired property.
To compute a nontrivial linear combination of the all zero vector using points from A 0 , we
solve the following linear program:

31 + 2 + 23 + 24 + 5 = 0
1 + 22 + 44 + 5 =0
21 + 52 + 3 + 34 + 5 = 0
1 + 2 + 3 + 4 + 5 =0

Using standard methods, e.g. gaussian elimination, one can compute the solution set as:

3 1
S = (a, 0, a, a, a) : a R .
2 2

We take the solution (1, 0, 32 , 21 , 1) S which gives the nontrivial linear combination

x1 x2 3 x3 1 x4 x5
0= +0 + .
1 1 2 1 2 1 1

As shown in the proof of Radons lemma, the sets A 1 := {x 1 , x 2 , x 5 } and A 2 := {x 3 , x 4 } have the
required property, i.e conv(A 1 ) conv(A 2 ) 6= ;.
As a certificate the proof gives that

2
1 1 3 1
v = x1 + x5 = x3 + x4 = 1
2 2 4 4
1.5

is contained in both conv A 1 and conv A 2 .

2
Exercise 3
Consider the vectors

3 1 2 2 1
x1 = 1 , x2 = 2 , x3 = 0 , x4 = 4 , x5 = 1 .

2 5 1 3 1

The vector
15
v = x 1 + 3x 2 + 2x 3 + x 4 + 3x 5 = 14
25
is a conic combination of the x i .
Write v as a conic combination using only three vectors of the x i .
Hint: Recall the proof of Carathodorys theorem
Solution
Let X = {x 1 , . . . , x n }. Observe that v cone(X ). Since x 1 , . . . , x 5 R3 , Carathodorys theorem
states that we can write v as a conic combination using at most three vectors of X .
How to compute this conic combination? Recall the proof of Carathodorys theorem.
The number of vectors in the conic combination v = 5i =1 i x i can be reduced by one with
P

the following method: Compute a nontrivial linear combination of the all zero vector, i.e.
compute 1 , . . . , 5 R, not all of them zero such that 5i =1 i x i = 0 holds.
P
P5
Thus v = i =1 (i i )x i for each > 0. As described in the proof, one can find an such
that i i 0 for each i = 1, . . . , 5 and i i = 0 for at least one i . Thus we get a new
conic combination of v using one vector less than before.
We now applying the idea to the exercise. We first compute a nontrivial linear combination
of the all zero vector by solving the following system of linear equations:

31 + 2 + 23 + 24 + 5 = 0
1 + 22 + 44 + 5 =0
21 + 52 + 3 + 34 + 5 = 0.

Using standard methods, e.g. gaussian elimination, one can compute the solution set as:

S = {(4a b, 0, 5a + b, a, b) : a, b R}

We take the solution (5, 0, 6, 1, 1) S which gives a nontrivial linear combination, i.e. 0 =
5x 1 + 6x 3 + x 4 + x 5 .
What is the maximal such that

v = (1 + 5)x 1 + 3x 2 + (2 6)x 3 + (1 )x 4 + (3 )x 5

is a conic combination? Each coefficient has to be nonnegative, thus observe that

1
=
3
is the maximum. We get the new conic combination
8 2 8
v = x 1 + 3x 2 + 0x 3 + x 4 + x 5 .
3 3 3

3
Observe that since the coefficient of x 3 is zero, we can remove it from the conic combination.
We need to remove one more vector to get a conic combination using only three vectors.
Again we compute a nontrivial linear combination of the all zero vector using the remaining
vectors (x 1 , x 2 , x 4 , x 5 ):

31 + 2 + 24 + 5 =0
1 + 22 + 44 + 5 =0
21 + 52 + 34 + 5 = 0.

We compute the solution set which is

S 0 = {(a, 0, a, 5a) : a R}

We take the solution (1, 0, 1, 5) S 0 which gives a nontrivial linear combination, i.e. 0 =
x 1 x 3 + 5x 5 .
What is the maximal such that
8 2 8
v = ( + )x 1 + 3x 2 + ( + )x 4 + ( 5)x 5
3 3 3
8
is a conic combination? It is given by = 15 .
The new conic combination is
8 8 2 8 8 8
v = ( + )x 1 + 3x 2 + ( + )x 4 + ( 5 )x 5
3 15 3 15 3 15
16 6
= x 1 + 3x 2 + x 4 + 0x 5 .
5 5
Since the coefficient of x 5 is zero, we can remove it and obtain the desired convex combina-
tion of v using only three vectors.
Exercise 4
Show that a basic solution can be associated to two different bases, i.e. give an example of
a solution x to a linear program min{c T x : Ax = b, x 0} such that there are two bases A B
and A B 0 with A B x B = b, A B 0 x B 0 = b and x (i ) = 0 i { j = 1, . . . , n : j B B 0 }.
Solution
Consider the linear program
Pn
min i =1 x i
Ax =b
x 0

Where
1 0 1 1
A := and b = .
0 1 1 0
Let B = {1, 2} and B 0 = {1, 3}.
Set x := (1, 0, 0)T . Observe that x has all desired properties.

4
Exercise 5
Recall the naive algorithm given in the lecture to solve a linear program by generating all
basic solutions. Consider linear programs of the form

min{c T x : Ax = b, x 0},

where A Qmn , b Qm and c Qn .

Assume that you have a computer that for every subset J {1, . . . , n} can check whether A J
T
is a basis, compute x = A 1
J b, check whether x 0 and compute c x in 1 msec.
If n = 2m, what is the largest m such that this computer can calculate an optimal solution
of the linear program using the naive algorithm in

1. one minute

2. one day

3. one year (365 days)

Solution
The time needed is given by the number of sets we have to test: The naive algorithm consid-
ers each
of {1, . . . , n} of cardinality m. For each such set 1 msec is needed.
subset Since
there
n 2m 2m
are = such sets, the naive algorithm needs a running time of msec.
m m m

18 20
One minute has 60,000 msec. = 48, 620 and = 184, 756. Thus m = 9 is the
9 10
largest m that can be processed in lessthan one minute.
28 30
One day has 86,400,000 msec. = 40, 116, 600 and = 155, 117, 520. Thus m = 14
14 15
is the largest m that can be processed in lessthan one day.
36 38
One year has 31,536,000,000 msec. = 9, 075, 135, 300 and = 35, 345, 263, 800.
18 19
Thus m = 18 is the largest m that can be processed in less than one year.