Numerical Optimization
Linear Programming - II
Shirish Shevade
Computer Science and Automation
Indian Institute of Science
Bangalore 560 012, India.
NPTEL Course on Numerical Optimization
Shirish Shevade
Numerical Optimization
LP in Standard Form:
min
cT x
s.t. Ax = b
x0
where A Rmn and rank(A) = m.
Shirish Shevade
Numerical Optimization
Let x be a nondegenerate basic feasible solution corresponding
to the basic variable set B and non-basic variable set N.
Let B denote the basis matrix.
Ax = b
BxB + NxN = b
xB = B1 b B1 NxN
Particular Solution: xB = B1 b and xN = 0
Objective Function =
=
=
=
cT x
cTB xB + cTN xN
cTB B1 b cTB B1 NxN + cTN xN
z + cTB xB + cTN xN
where cTB = 0T and cTN = cTN cTB B1 N are the relative cost
factors corresponding to the basis matrix B and z denotes the
current objective function value.
Shirish Shevade
Numerical Optimization
LP in Standard Form:
min
cT x
s.t. Ax = b
x0
where A Rmn and rank(A) = m.
Convex Programming Problem
Assumption:
Feasible set is non-empty
Slaters condition is satisfied
First order KKT conditions are necessary and sufficient at
optimality
Shirish Shevade
Numerical Optimization
LP in Standard Form:
min
cT x
s.t. Ax = b
x0
where A Rmn and rank(A) = m.
Define the Lagrangian function:
L(x, , ) = cT x + T (b Ax) T x
First Order KKT Conditions at optimality:
Primal Feasibility: Ax = b, x 0
x L(x, , ) = 0 AT + = c
Complementary Slackness Condition: i xi = 0 i
Non-negativity: i 0 i
Shirish Shevade
Numerical Optimization
Let x be a nondegenerate basic feasible solution corresponding
to the basic variable set B and non-basic variable set N.
x = (xB xN )T where xB > 0 and xN = 0.
At optimal (x, , ),
B = 0 and N 0.
c = AT + . That is,
T
cB
B
B
cB = BT + B
=
cN
N
NT
cN = NT + N
B = 0 cB = BT = BT
N 0 requires that
cB
N = cN (B1 N)T cB 0
The current basic feasible solution x is not optimal if there
exists xq N such that q < 0.
Shirish Shevade
Numerical Optimization
x is feasible Ax = b, x 0.
BxB + NxN = b xB = B1 b B1 NxN
cT x
cTB xB + cTN xN
cTB (B1 b B1 NxN ) + cTN xN
cTB B1 b + (cTN cTB B1 N) xN
|
{z
}
T
N
T 1
= cB B b
( xN = 0)
Objective Function at x =
=
=
=
Suppose there exists a non-basic variable xq N (xq = 0) such
that q < 0.
Objective Function at x = cTB B1 b + q xq
The objective function can be decreased if xq is changed from
0 to some positive value (by making xq a basic variable).
Shirish Shevade
Numerical Optimization
Suppose xq is made a basic variable. Therefore, some existing
basic variable xp ( B) needs to be made non-basic.
Ax = b
BxB + NxN = b
xB = B1 b B1 NxN
x1
b1
..
..
.
.
1
xp = b
p (B N),q xq
.
.
..
..
xm
bm
Note: b > 0
How to choose xp ?
Require (B1 N)p,q > 0 so that xp can decrease if xq increases.
Shirish Shevade
Numerical Optimization
x1
b1
..
..
.
.
xp = bp (B1 N),q xq
.
.
..
..
xm
bm
How to choose the basic variable xp to be made non-basic?
q = {xj B : (B1 N)jq > 0}.
Let B
xq = min
q
xj B
bj
(B1 N) jq
bp
(B1 N) pq
q = , the problem is unbounded.
If B
Shirish Shevade
Numerical Optimization
Summary of steps:
Given a basic feasible solution, (xB xN )T w.r.t. B.
If N = cN (B1 N)T cB 0
1
Choose xq N such that q < 0
q = {xj B : (B1 N)jq > 0}.
2
B
q 6= , find the basic variable xp to be made non-basic.
3
If B
xq = min
q
xj B
4
5
bj
(B1 N) jq
bp
(B1 N) pq
xi = bi (B1 N)iq xq , xi B
Swap xp and xq from the sets B and N
Geometrical Interpretation of Steps 1-5
Moving from one extreme point of the feasible set to an
adjacent extreme point so that the objective function value
decreases.
Shirish Shevade
Numerical Optimization
Example:
min 3x1 x2
s.t. x1 + x2 2
x1 1
x1 , x2 0
Shirish Shevade
Numerical Optimization
LP in Standard Form:
min
cT x
s.t. Ax = b
x0
where A Rmn and rank(A) = m.
Given a basis matrix B w.r.t. the basis vector set B,
Basic Feasible Solution: xB = B1 b, xN = 0
Objective function = cTB B1 b
Relative cost factors, N = cN (B1 N)T cB
Shirish Shevade
Numerical Optimization
LP in Standard Form:
min
cT x
s.t. Ax = b
x0
where A Rmn and rank(A) = m.
Basic Variables Nonbasic Variables RHS
xB
B
N
b
cTB
cTN
0
I
0T
I
cTB
B1 b
0
B1 N
cTN
B1 N
cTN cTB B1 N
Shirish Shevade
B1 b
- cTB B1 b
Numerical Optimization
Example:
min 3x1 x2
s.t. x1 + x2 2
x1 1
x1 , x2 0
Given problem in the standard form:
min
3x1 x2
s.t. x1 + x2 + x3 = 2
x1 + x4 = 1
x1 , x2 , x3 , x4 0
Initial Basic Feasible Solution:
xB = (x3 , x4 )T = (2, 1)T , xN = (x1 , x2 )T = (0, 0)T
Shirish Shevade
Numerical Optimization
min
3x1 x2
s.t. x1 + x2 + x3 = 2
x1 + x4 = 1
x1 , x2 , x3 , x4 0
Initial Basic Feasible Solution:
xB = (x3 , x4 )T = (2, 1)T , xN = (x1 , x2 )T = (0, 0)T
Initial Tableau:
x1 x2 x3 x4 RHS
1 1 1 0
2
1 0 0 1
1
-3 -1 0 0
0
Shirish Shevade
Numerical Optimization
Initial Basic Feasible Solution:
xB = (x3 , x4 )T = (2, 1)T , xN = (x1 , x2 )T = (0, 0)T
Current Objective function = 0
Shirish Shevade
Numerical Optimization
x1 x2 x3
1 1 1
1 0 0
-3 -1 0
x4
0
1
0
RHS
2
1
0
Incoming basic variable: x1
Shirish Shevade
Numerical Optimization
x1 x2
1 1
1 0
-3 -1
Incoming nonbasic variable:
x1 x2
1 1
1 0
-3 -1
Outgoing basic variable: x4
x1 x2
0 1
1 0
0 -1
x3 x4
1 0
0 1
0 0
x1
RHS
2
1
0
x3
1
0
0
x4
0
1
0
RHS
2
1
0
x3
1
0
0
x4
-1
1
3
RHS
1
1
3
Shirish Shevade
Numerical Optimization
x1
0
1
0
x2
1
0
-1
x3
1
0
0
x4
-1
1
3
RHS
1
1
3
Current Basic Feasible Solution:
(x1 , x3 )T = (1, 1)T , (x2 , x4 )T = (0, 0)T
Current Objective function = 3
Shirish Shevade
Numerical Optimization
Current Basic Feasible Solution:
(x1 , x3 )T = (1, 1)T , (x2 , x4 )T = (0, 0)T
Current Objective function = 3
Shirish Shevade
Numerical Optimization
x1
0
1
0
x2 x3
1 1
0 0
-1 0
x4
-1
1
3
RHS
1
1
3
Incoming nonbasic variable: x2
Shirish Shevade
Numerical Optimization
x1 x2
0 1
1 0
0 -1
Incoming nonbasic variable:
x1 x2
0 1
1 0
0 -1
Outgoing basic variable: x3
x1 x2
0 1
1 0
0 0
x3 x4
1 -1
0 1
0 3
x2
RHS
1
1
3
x3
1
0
0
x4
-1
1
3
RHS
1
1
3
x3
1
0
1
x4
-1
1
2
RHS
1
1
4
Shirish Shevade
Numerical Optimization
Current Basic Feasible Solution:
(x1 , x2 )T = (1, 1)T , (x3 , x4 )T = (0, 0)T
Current Objective function = 4
Shirish Shevade
Numerical Optimization
x1
0
1
0
x2
1
0
0
x3
1
0
1
x4
-1
1
2
RHS
1
1
4
Current Basic Feasible
Solution: (x1 , x2 )T =
(1, 1)T , (x3 , x4 )T = (0, 0)T
Current Objective function =
4
N = (1, 2)T > 0T
Optimal Solution:
x = (1, 1)T , cT x = 4
Shirish Shevade
Numerical Optimization
Initial Basic Feasible Solution:
xB = (x3 , x4 )T = (2, 1)T , xN = (x1 , x2 )T = (0, 0)T
Current Objective function = 0
Shirish Shevade
Numerical Optimization
x1 x2
1 1
1 0
-3 -1
Incoming basic variable: x2
x3
1
0
0
Shirish Shevade
x4
0
1
0
RHS
2
1
0
Numerical Optimization
LP in Standard Form:
min
cT x
s.t. Ax = b
x0
where A Rmn and rank(A) = m.
Simplex Algorithm to solve an LP [Dantzig]1
G. B. Dantzig, Linear Programming and Extensions, Princeton
University Press, 1963
Shirish Shevade
Numerical Optimization
Simplex Algorithm (to solve an LP in Standard Form)
(1) Get initial basis matrix B, basic feasible solution,
xB = B1 b, xN = 0, set Unbounded = FALSE.
(2) N = cN (B1 N)T cB
(3) while (Unbounded == FALSE and xj N 3 j < 0)
(a) Select a non-basic variable xq N such that q < 0
q = {xj B : (B1 N)jq > 0}.
(b) B
q ==
(c) if B
Unbounded = TRUE
else
(i) xq = minxj Bq
bj
bp
(B1 N) jq
(B1 N) pq
(ii) xi =
bi (B N)iq xq , xi B
(iii) Swap xp and xq between B and N, update B and
1
endif
endwhile
Output : x = (xB , xN )T if Unbounded == FALSE
Shirish Shevade
Numerical Optimization
Remarks:
How to select a non-basic variable xq N in Step 3(a) of
the Simplex Algorithm?
xq = argmin j
xj N:j <0
If the basic feasible solution is nondegenerate at each
iteration, then the Simplex Algorithm terminates in a finite
number of iterations.
Objective Function = cTB xB + cTN xN
= cTB B1 b cTB B1 NxN + cTN xN
= z + cTB xB + cTN xN
where cTB = 0T , cTN = cTN cTB B1 N and z denotes the
current objective function value. If the basic feasible
solution is nondegenerate, then the objective function
decreases in each iteration. Since the number of basic
feasible solutions is finite, the algorithm has finite
convergence.
Shirish Shevade
Numerical Optimization
How to get initial basic feasible solution?
(a) Case I: Given constraints
Ax b
x0
(with b 0), can be written as
x
Ax + y = b
(A I)
=b
y
x, y 0
x, y 0.
Use Basic feasible solution, (x, y)T = (0, b)T , to solve
min cT x + 0T y
s.t. Ax + y = b
x, y 0
Shirish Shevade
Numerical Optimization
(b) Case 2: Given constraints (C2)
Ax = b
x 0,
(with b 0),
solve the artificial linear program (ALP)
min
1T y
s.t. Ax + y = b
x, y 0.
to get the initial basic feasible solution, x 0 (if it exists).
If there exists x that satisfies constraints C2, then ALP has
the optimal objective function value of 0 with y = 0 and
x 0.
If C2 has no feasible solution, then the optimal objective
function value of ALP is greater than 0.
Shirish Shevade
Numerical Optimization
(c) Case 3: Given constraints
Ax b
x0
(with b 0), can be written as
x
Ax z + y = b
(A I I) y = b
z
x, y, z 0
x, y, z 0.
Solve an artificial linear program (similar to Case 2)
Shirish Shevade
Numerical Optimization
Two Phase Method
Given the linear program (SLP):
min
cT x
s.t. Ax = b
x0
where b 0.
Phase I: Introduce artificial variables and solve the
artificial linear program with initial basic feasible solution,
(x = 0, y = b), to get a basic feasible solution for SLP or
conclude that it does not exist.
Phase II: Get the initial basic feasible solution (if it exists)
from Phase I to solve SLP.
Shirish Shevade
Numerical Optimization
Example:
min
s.t.
2x1 x2
x1 + x2 3
x1 + x2 1
x1 , x2 0
Shirish Shevade
Numerical Optimization
min
s.t.
Example:
2x1 x2
x1 + x2 3
x1 + x2 1
x1 , x2 0
Phase I: Introducing artificial variables, the constraints become
x1 + x2 + x3 = 3
x1 + x2 x4 + x5 = 1
x1 , x2 , x3 , x4 , x5 0
Therefore, the artificial linear program
min
s.t.
x3 + x5
x1 + x2 + x3 = 3
x1 + x2 x4 + x5 = 1
x1 , x2 , x3 , x4 , x5 0
has x3 = 3, x5 = 1, x1 = x2 = x4 = 0 as its initial basic feasible
solution.
Shirish Shevade
Numerical Optimization
min
s.t.
Initial Tableau:
x1
1
-1
0
x3 + x5
x1 + x2 + x3 = 3
x1 + x2 x4 + x5 = 1
x1 , x2 , x3 , x4 , x5 0
x2
1
1
0
x3
1
0
1
x4 x5
0 0
-1 1
0 1
RHS
3
1
0
Making the relative costs of basic variables 0,
x1 x2 x3 x4 x5 RHS
1 1 1 0 0
3
-1 1 0 -1 1
1
0 -2 0 1 0
-4
Shirish Shevade
Numerical Optimization
x1 x2 x3
1 1 1
-1 1 0
0 -2 0
x4 x5
0 0
-1 1
1 0
Incoming non-basic variable: x2
x1 x2 x3 x4
1 1 1 0
-1 1 0 -1
0 -2 0 1
Outgoing basic variable: x5
x1 x2 x3
2 0 1
-1 1 0
-2 0 0
Shirish Shevade
RHS
3
1
-4
x5
0
1
0
RHS
3
1
-4
x4 x5
1 -1
-1 1
-1 2
RHS
2
1
-2
Numerical Optimization
x1 x2 x3 x4 x5
2 0 1 1 -1
-1 1 0 -1 1
-2 0 0 -1 2
Incoming non-basic variable: x1
x1 x2 x3 x4 x5
2 0 1 1 -1
-1 1 0 -1 1
-2 0 0 -1 2
Outgoing basic variable: x3
x1 x2 x3
1 0 1
2
0 1 1
2
0 0 1
x4
1
2
21
0
x5
12
1
2
RHS
2
1
-2
RHS
2
1
-2
RHS
1
2
0
Basic feasible solution: x1 = 1, x2 = 2, x3 = x4 = 0.
Shirish Shevade
Numerical Optimization
x1 x2 x3 x4
x5 RHS
1
1
1 0
12
1
2
2
1
0 1 1 1
2
2
2
2
0 0 1
0
1
0
Basic feasible solution: x1 = 1, x2 = 2, x3 = x4
Shirish Shevade
Numerical Optimization
= 0.
Phase II: For the given problem,
min
s.t.
2x1 x2
x1 + x2 + x3 = 3
x1 + x2 x4 = 1
x1 , x2 , x3 , x4 0
Initial Tableau:
x1
1
0
2
x2
0
1
-1
x3
x4
1
2
1
2
1
2
21
0
RHS
1
2
0
Making the relative costs of basis variables 0,
x1 x2 x3
x4 RHS
1
1
1 0
1
2
2
1
1
0 1
2
2
2
0 0 12 23
0
Shirish Shevade
Numerical Optimization
12
32
RHS
1
2
0
Incoming non-basic variable: x4
x1 x2 x3
x4
1
1
1 0
2
2
1
1
0 1
2
2
1
0 0 2 32
RHS
1
2
0
x1
1
0
0
x2
0
1
0
Outgoing basic variable: x1
x1 x2
2 0
1 1
3 0
x3
x4
1
2
1
2
1
2
12
x3
1
1
1
x4
1
0
0
RHS
2
3
3
x = (x1 , x2 )T = (0, 3)T , cT x = 3
Shirish Shevade
Numerical Optimization
x1
2
1
3
x2
0
1
0
x3
1
1
1
Shirish Shevade
x4
1
0
0
RHS
2
3
3
Numerical Optimization
Example:
min 4x1 2x2
s.t. 3x1 2x2 4
2x1 + x2 = 2
x1 , x2 0
Phase I: Introducing artificial variables, the constraints become
3x1 2x2 x3 + x4 = 4
2x1 + x2 + x5 = 2
x1 , x2 , x3 , x4 , x5 0
Therefore, the artificial linear program
min
x4 + x5
s.t. 3x1 2x2 x3 + x4 = 4
2x1 + x2 + x5 = 2
x1 , x2 , x3 , x4 , x5 0
has x4 = 4, x5 = 2, x1 = x2 = x3 = 0 as its initial basic feasible
solution.
Shirish Shevade
Numerical Optimization
min
x4 + x5
s.t. 3x1 2x2 x3 + x4 = 4
2x1 + x2 + x5 = 2
x1 , x2 , x3 , x4 , x5 0
Initial Tableau:
x1 x2
3 -2
-2 1
0 0
x3
-1
0
0
x4
1
0
1
x5
0
1
1
RHS
4
2
0
Making the relative costs of basic variables 0,
x1 x2 x3 x4 x5 RHS
3 -2 -1 1 0
4
-2 1 0 0 1
2
-1 1 1 0 0
-6
Shirish Shevade
Numerical Optimization
x1 x2 x3 x4
3 -2 -1 1
-2 1 0 0
-1 1 1 0
Incoming non-basic variable: x1
x1 x2 x3 x4
3 -2 -1 1
-2 1 0 0
-1 1 1 0
x5
0
1
0
RHS
4
2
-6
x5
0
1
0
RHS
4
2
-6
Outgoing basic variable: x4
x1 x2 x3
1 2 1
3
3
0 1 2
3
3
1
2
0
3
3
x5
0
1
0
RHS
x4
1
3
2
3
1
3
4
3
14
3
- 143
The given problem has no feasible solution.
Shirish Shevade
Numerical Optimization
min 4x1 2x2
s.t. 3x1 2x2 4
2x1 + x2 = 2
x1 , x2 0
Shirish Shevade
Numerical Optimization
Cycling
Possible in the presence of degenerate basic feasible
solution
Consider the example given by Beale2
min
s.t.
34 x1 + 20x2 12 x3 + 6x4
1
4 x1
1
2 x1
8x2 x3 + 9x4 = 0
12x2 12 x3 + 3x4 = 0
x3 + x5 = 1
xj 0 j
Blands rule to avoid cycling: Arrange the variables in
some ordered sequence. Among all the candidate variables
to enter (or to leave) the basis, choose the variable with the
smallest index
2
Beale E.M.L., Cycling in the dual simplex algorithm, Naval Research
Logistics Quarterly, 2(4), pp. 269-276, 1955.
Shirish Shevade
Numerical Optimization