Lecture 10: Linear Programming Duality and Sensitivity
Lecture 10: Linear Programming Duality and Sensitivity
0-0
' $
1
maximize z = cT x (1)
subject to Ax ≤ b,
x ≥ 0n
and
minimize w = bT y (2)
subject to AT y ≥ c,
y ≥ 0m
& %
' $
2
minimize z = cT x (P)
subject to Ax = b,
x ≥ 0n
and
maximize w = bT y (D)
subject to AT y ≤ c,
y free
& %
' $
3
& %
' $
4
& %
' $
5
& %
' $
7
& %
' $
8
& %
' $
9
& %
vector and see that it can later be eliminated
' $
10
& %
' $
11
& %
' $
12
The feasibility of y T = cT
B B −1
is not fulfilled until we
reach an optimal BFS. How is the incoming criterion
related to this? We introduce as an incoming variable a
variable which has the best reduced cost. Since the
reduced cost measures the dual feasibility of y, this
means that we select the most violated dual constraint;
at the new BFS, that constraint is then satisfied (since
the reduced cost then is zero). The Simplex method
hence works to try to satisfy dual feasibility by forcing
a move such that the most violated dual constraint
becomes satisfied!
& %
' $
13
& %
Proof.
' $
14
& %
' $
16
∗
Sensitivity analysis, I: Shadow prices are
derivatives of a convex function!
Suppose an optimal BFS is non-degenerate. Then,
cT x∗ = bT y ∗ = cT
B B −1
b varies linearly as a function of
b around its given value
Non-degeneracy also implies that y ∗ is unique. Why?
Perturbation function b 7→ v(b) given by
v(b) = min cT x = max bT y = max bT y k
k∈K
s.t. Ax = b, s.t. AT y ≤ c
x ≥ 0n
K: set of DFS. v a piece-wise linear, convex function
& %
' $
17
& %
' $
18
∗
Sensitivity analysis, II: Perturbations in data
How to find a new optimum through re-optimization
when data has changed
If an element of c changes, then the old BFS is feasible
but may not be optimal. Check the new value of the
reduced cost vector c̃ and change the basis if some sign
has changed
& %
' $
19
∗
Decentralized planning
Consider the following profit maximization problem:
m
X
maximize z = pT x = pT
i xi ,
i=1
B1 b
1
x1
B2 b2
x2
.. ..
s.t. . ·
.. ≤ . ,
.
Bm bm
xm
C c
x i ≥ 0 ni ,
& %
i = 1, . . . , m,
' $
21
& %
' $
22
The units also use limited resources that are the same
The resource constraint is difficult as well as unwanted
to enforce directly, because it would make it a
centralized planning process
We want the units to maximize their own profits
individually
But we must also make sure that they do not violate
the resource constraints Cx ≤ c
Pm
(This constraint is typically of the form i=1 C i xi ≤ c)
How?
ANSWER: Solve the LP dual problem!
& %
' $
23
& %