Dual Opt
Uploaded by
Shahzad AshrafDual Opt
Uploaded by
Shahzad AshrafRelations between Primal and Dual
If the primal problem is
Maximize ct x subject to Ax = b, x ≥ 0
then the dual is
Minimize bt y subject to At y ≥ c (and y unrestricted)
Easy fact:
If x is feasible for the primal, and y is feasible for the dual, then
ct x ≤ bt y
So (primal optimal) ≤ (dual optimal) (Weak Duality Theorem)
Much less easy fact: (Strong Duality Theorem)
If one of the primal and the dual have finite optima, they both have and
(primal optimal) = (dual optimal)
1
The “Inverse Matrix”
In the initial simplex tableau, there’s an identity matrix. At a later simplex
tableau, the “inverse matrix” is the matrix occupying the same space as
that original identity matrix.
The inverse matrix conveys all information about the current state of the
algorithm, as we will see.
2
3
Computing dual values from Inverse Matrix
If we have reached the optimal primal tableau, these methods give the
optimal dual values; at earlier iterations, they give a certain “dual” of the
current basic feasible solution
Method 1:
Row vector of dual values = Row vector of original objective values of
current basic variables (listed in order they appear along basic column of
current tableau) X current inverse
Method 2: (see textbook for this)
Write wi for initial basic variable in row i.
Value of dual variable yi = current z-row coefficient of wi + original
objective coefficient of wi
4
Example — Primal problem
Minimize 4x1 + 2x2 − x3
subject to
x1 + x2 + 2x3 ≥ 3
2x1 − 2x2 + 4x3 ≤ 5
x1 , x2 , x3 ≥ 0.
In standard form:
Minimize 4x1 + 2x2 − x3 + 0s1 + M R + 0s2
subject to
x1 + x2 + 2x3 − s1 + R = 3
2x1 − 2x2 + 4x3 + s2 = 5
x1 , x2 , x3 , s1 , R, s2 ≥ 0.
5
Example — The dual
Maximize 3y1 + 5y2
subject to
y1 + 2y2 ≤ 4
y1 − 2y2 ≤ 2
2y1 + 4y2 ≤ −1
−y1 ≤ 0, i.e. y1 ≥ 0
y1 ≤ M
y2 ≤ 0
Note that the addition of the artificial variable to the primal adds a new
constraint to the dual: y1 ≤ M . But since we imagine M to be very large,
this effectively puts no new constraint on y1 . For convenience, we’ll take
M = 100.
6
Example — initial tableau
x1 x2 x3 s1 R s2 soln
z-row 96 98 201 −100 0 0 300
R 1 1 2 −1 1 0 3
s2 2 −2 4 0 0 1 5
Inverse matrix is bolded
Current solution: R = 3, s2 = 5, z = 300. Feasible, but not optimal
Current “dual solution”:
1 0
(y1 y2 ) = (100 0) = (100 0)
0 1
i.e. y1 = 100, y2 = 0, zdual = 300
We will see later that this is “optimal but not feasible”.
7
Example — first iteration
x1 x2 x3 s1 R s2 soln
z-row −4.5 198.5 0 −100 0 −50.25 48.75
R 0 2 0 −1 1 −.5 .5
x3 .5 −.5 1 0 0 .25 1.25
Inverse matrix is bolded
Current solution: R = .5, x3 = .25, z = 48.75. Feasible, but not optimal
Current “dual solution”:
1 −.5
(y1 y2 ) = (100 − 1) = (100 − 50.25)
0 .25
i.e. y1 = 100, y2 = −50.25, zdual = 48.75
Again, “optimal but not feasible”.
8
Example — optimal tableau
x1 x2 x3 s1 R s2 soln
z-row −4.5 0 0 −.75 −99.25 −.625 −.875
x2 0 1 0 −.5 .5 −.25 .25
x3 .5 0 1 −.25 .25 .125 1.375
Inverse matrix is bolded
Optimal solution: x2 = .25, x3 = 1.375, z = −.875.
Optimal dual solution:
.5 −.25
(y1 y2 ) = (2 − 1) = (.75 − .625)
.25 .125
i.e. y1 = .75, y2 = −.625, zdual = −.875
This is readily checked to be feasible and optimal for dual
9
Getting whole tableau from Inverse (and initial data)
Constraint columns:
New constraint column = current inverse X original constraint column
Objective coefficients:
Objective coefficient (z-row entry) of variable xj = Left hand side of jth
dual constraint (evaluated at current “dual solution”) - right hand side of
jth dual constraint
10
Example — first iteration of previous problem
x1 x2 x3 s1 R s2 soln
z-row −4.5 198.5 0 −100 0 −50.25 48.75
R 0 2 0 −1 1 −.5 .5
x3 .5 −.5 1 0 0 .25 1.25
We look at (bolded) x2 constraint column and objective entry
1 −.5 1 2
=
0 .25 −2 −.5
(Inverse X original x2 column = new x2 column)
Coefficient of x2 in z-row is computed by
(y1 − 2y2 ) − 2 = ([100] − 2[−50.25]) − 2 = 198.5
using values of y1 , y2 computed earlier.
11
