21
updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Embed Size (px)

Citation preview

Page 1: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

updated 21 April2008

Linear Programs with Totally Unimodular Matrices

Page 2: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Basic Feasible Solutions

0,

)2(4595

)1(6s.t.

85max

yx

yx

yx

yxz

0,,,

4595

6s.t.

85max

21

2

1

ssyx

syx

syx

yxz

Standard Form

slide 2

Page 3: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Basic Feasible Solutions

0,,,

4595

6s.t.

85max

21

2

1

ssyx

syx

syx

yxz

Solution Basic Variables Non-Basic Variables Intersection

BFS 1 x = 2.25 & y = 3.75 s1 = s2 = 0 (1) and (2)

BFS 2 x = 6 & s2 = 15 y = s1 = 0 (1) and x-axis

BFS 3 y = 5 & s1 = 1 x = s1 = 0 (2) and y-axis

BFS 4 s1 = 6 & s2 = 45 x = y = 0 x-axis and y-axis

slide 3

Page 4: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Vector-Matrix Representation

45

6

1095

011121

b

ssyx

A

45

6,

10

01,,

1

5,

09

11,,

15

6,

15

01,,

75.3

25.2,

95

11,,

121

11

12

1

bAAssB

bAAsyB

bAAsxB

bAAyxB

BB

BB

BB

BB

slide 4

Page 5: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Example MCNFP

5 1 4(1, 0,2)

3

2

0

-3

-2

(2, 0,2)

(4, 1,3)

(4, 0,3)

(3, 2,5)

slide 5

Page 6: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

LP for Example MCNFP

Min 3x12 + 2 x13 + x23 + 4 x24 + 4 x34 s.t. x12 + x13 = 5 {Node 1} x23 + x24 – x12 = -2 {Node 2}

x34 – x13 - x23 = 0 {Node 3} – x24 - x34 = -3 {Node 4}

2 x12 5, 0 x13 2, 0 x23 2, 1 x24 3,

0 x34 3,

slide 6

Page 7: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Matrix Representation of Flow Balance Constraints

3025

11000101100110100011

34

24

23

13

12

xxxxx

slide 7

Page 8: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Solving for a Basic Feasible Solution

3025

1100101001010011

34

24

13

12

xxxx

3025

1100101001010011

1

34

24

13

12

xxxx

slide 8

Page 9: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Cramer’s Rule

Use determinants to solve x=A-1b.

abaa

abaaB

AB

xnnnnn

nij

j

j

j

21

1112

,

Take the matrix A and replace column j with the vector b to form matrix Bj.

slide 9

Page 10: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Using Cramer’s Rule to Solve for x12

A

Bx

)2,1(

12

1100

1010

0101

0011

1103

1010

0102

0015

integer?an Is )2,1(B

?1Does Aslide 10

Page 11: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Total Unimodularity

• A square, integer matrix is unimodular if its determinant is 1 or -1.

• An integer matrix A is called totally unimodular (TU) if every square, nonsingular submatrix of A is unimodular.

TUTUNot 1100101001010011

1111

slide 11

Page 12: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Total Unimodularity

• A square, integer matrix is unimodular if its determinant is 1 or -1.

• An integer matrix A is called totally unimodular (TU) if every square, nonsingular submatrix of A is unimodular.

1100

1010

0101

0011

111

01

slide 12

Page 13: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Sufficient Conditions for TU

An integer matrix A is TU if1. All entries are -1, 0 or 12. At most two non-zero entries appear in any column3. The rows of A can be partitioned into two disjoint

sets M1 and M2 such that• If a column has two entries of the same sign, their rows are

in different sets.• If a column has two entries of different signs, their rows are

in the same set.

slide 13

Page 14: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

The Matrix of Flow Balance Constraints

3025

11000101100110100011

34

24

23

13

12

xxxxx

• Every column has exactly one +1 and exactly one -1.• This satisfies conditions 1 and 2.

• Let the row partition be M1 = {all rows} and M2 = {}.• This satisfies condition 3.

• Thus the flow balance constraint matrix is TU.

slide 14

Page 15: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Using Cramer’s Rule to Solve for x12

A

Bx

)2,1(

12

1100

1010

0101

0011

1103

1010

0102

0015

integer?an Is )2,1(B?1Does A Yes.

slide 15

Page 16: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Expansion by Minors: 4-by-4 Matrix

aaaaaaaaa

aaaaaaaaaa

a

aaaaaaaaa

aaaaaaaaaa

a

aaaaaaaaaaaaaaaa

aaaaaaaaaaaaaaaa

342414

332313

322212

41

442414

432313

422212

31

443414

433313

423212

21

443424

433323

423222

11

44342414

43332313

42322212

41312111

44434241

34333231

24232221

14131211

slide 16

Page 17: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Expansion by Minors: 3-by-3 Matrix

aaaaa

aaaaa

aaaaa

aaaaaaa

aaaaaaaa

aaaaaaaaa

3122322113

3123332112

3223332211

3231

222113

3331

232112

3332

232211

333231

232221

131211

slide 17

Page 18: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Using Cramer’s Rule to Solve for x12

A

Bx

)2,1(

12

1100

1010

0101

0011

1103

1010

0102

0015

integer?an Is )2,1(B

?1Does A Yes.slide 18

Page 19: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Using Cramer’s Rule to Solve for x12

1100101001013025

1103101001020015

• When we expand along minors, the determinants of the submatrices will be +1, -1, or 0.• Therefore, the determinant will be an integer: (5)(+1, -1, or 0) + (-2) (+1, -1, or 0) + 0 + (-3) (+1, -1, or 0).

slide 19

Page 20: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

Using Cramer’s Rule to Solve for x12

A

Bx

)2,1(

12

1100

1010

0101

0011

1103

1010

0102

0015

integer?an Is )2,1(B

?1Does A Yes.

Yes.

slide 20

Page 21: Updated 21 April2008 Linear Programs with Totally Unimodular Matrices

TU Theorems

• Matrix A is TU if and only if AT is TU.• Matrix A is TU if and only if [A, I] is TU.

– I is the identity matrix.

• If the constraint matrix for an IP is TU, then its LP relaxation has an integral optimal solution.

• The BFSs of an MCNF LP are integer valued.

slide 21