CO 250: Introduction to Optimization

Module 4: Duality Theory (Strong Duality)

University
c of Waterloo 1 / 11
 Recap: Weak Duality

Last lecture: we described a

method to construct the dual of a max (2, −1, 3)x (P)
general linear program.    
1 0 −1 ≤ 2
E.g.: consider the primal LP, (P), on s.t. 0 −2 1  x =  1 
the right – a max LP that falls in 1 1 0 ≥ −2
the first two columns of the table.
x1 ≥ 0, x2 ≤ 0, x3 free
−→ The dual of (P) is a min LP.

University
c of Waterloo 2 / 11
 Recap: Weak Duality

max (2, −1, 3)x (P) min (2, 1, −2)y (D)

       
1 0 −1 ≤ 2 1 0 1 ≥ 2
s.t. 0 −2 1  x =  1  s.t.  0 −2 1 y ≤ −1
1 1 0 ≥ −2 −1 1 0 = 3
x1 ≥ 0, x2 ≤ 0, x3 free y1 ≥ 0, y2 free, y3 ≤ 0

University
c of Waterloo 3 / 11
 Recap: Weak Duality

max (2, −1, 3)x (P) min (2, 1, −2)y (D)

       
1 0 −1 ≤ 2 1 0 1 ≥ 2
s.t. 0 −2 1  x =  1  s.t.  0 −2 1 y ≤ −1
1 1 0 ≥ −2 −1 1 0 = 3
x1 ≥ 0, x2 ≤ 0, x3 free y1 ≥ 0, y2 free, y3 ≤ 0

Weak Duality Theorem

if x̄ is feasible for (P) and ȳ is feasible for (D),

=⇒ cT x̄ ≤ bT ȳ
If cT x̄ = bT ȳ, then both x̄ and ȳ are optimal.

University
c of Waterloo 4 / 11
 This Lecture: Strong Duality

Question
Can we always find feasible solutions x̄ and ȳ to a primal-dual pair,
(Pmax ), (Pmin ), such that cT x̄ = bT ȳ?

Strong Duality Theorem

If (Pmax ) has an optimal solution x̄, then (Pmin ) has an optimal
solution ȳ such that cT x̄ = bT ȳ.

University
c of Waterloo 5 / 11
 Strong Duality – for LPs in SEF

Let us prove the Strong Duality Theorem in

the special case where (P) is in SEF. max cT x (P)
s.t. Ax = b
Let’s assume (P) has an optimal solution. x≥0
−→ 2-Phase Simplex terminates with an
optimal basis B (Why?)
min bT y (D)
We can rewrite (P) for basis B: T
s.t. A y ≥ c
max z = ȳ T b + c̄T x (P’)
s.t. xB + A−1
B AN x N = A−1
B b
where:
x≥0 ȳ = A−T
B cB
c̄T = cT − ȳ T A
Thus, x̄N = 0 and x̄B = A−1
B b

University
c of Waterloo 6 / 11
 Strong Duality – for LPs in SEF

We can rewrite (P) for basis B:

max cT x (P)
max z = ȳ T b + c̄T x (P’)
s.t. Ax = b
s.t. xB + A−1 −1
B AN x N = AB b x≥0
x≥0

Thus, x̄N = 0 and x̄B = A−1

B b
min bT y (D)
Recall that (P) and (P’) are equivalent! T
s.t. A y ≥ c
−→ x̄ has same value in (P) and (P’)
where:
cT x̄ = ȳ T b + c̄T x̄
ȳ = A−T
B cB
= ȳ T b + c̄TN x̄N
c̄T = cT − ȳ T A
= bT ȳ

Goal: Show that ȳ is dual feasible.

University
c of Waterloo 7 / 11
 Strong Duality – for LPs in SEF

We can rewrite (P) for basis B:

max cT x (P)
max z = ȳ T b + c̄T x (P’) s.t. Ax = b
s.t. xB + A−1 −1
B AN x N = AB b x≥0
x≥0

Thus, x̄B = A−1B b and x̄N = 0 and min bT y (D)

cT x̄ = bT ȳ.
T
s.t. A y ≥ c
Note that B is an optimal basis −→ c̄ ≤ 0

−→ cT − ȳ T A ≤ 0 where:
ȳ = A−T
B cB
T
Equivalently, A ȳ ≥ c, c̄T = cT − ȳ T A
meaning ȳ is dual feasible!

University
c of Waterloo 8 / 11
 Strong Duality Theorem

Strong Duality Theorem

Let (P) and (D) be a primal-dual pair of LPs. If (P) has an optimal
solution, then (D) has one, and their objective values equal.

Note: (P) is feasible and (D) is feasible −→ (P) cannot be unbounded

Fundamental Theorem of LP −→ (P) has an optimal solution.

Subtly different version via previous results:

Strong Duality Theorem – Feasibility Version

Let (P) and (D) be primal-dual pair of LPs. If both are feasible, then
both have optimal solutions of the same objective value.

University
c of Waterloo 9 / 11
 Possible Outcomes of Primal-Dual Pair (P), (D)

(D)(P) optimal solution unbounded infeasible

optimal solution possible 1 impossible
2 impossible 3
unbounded impossible 4 impossible
5 possible 6
infeasible impossible 7 possible 8 possible 9

• ,
1 ,
6 and
8 many examples exist max cT x (P)
•
2 follows directly from Weak Duality as s.t. Ax = b
follows: x≥0
Suppose, for a contradiction, that (D) has an
optimal solution ȳ.
min bT y (D)
cT x̄ ≤ bT ȳ for all feasible primal solutions x̄
T
by Weak Duality −→ (P) is bounded! s.t. A y ≥ c
Similar arguments apply to 4 and 5
• ,
3 7 follow directly from Strong Duality
• I’ll leave
9 for you to do as an exercise!

University
c of Waterloo 10 / 11
 Recap

Strong Duality Theorem

Let (P) and (D) be a primal-dual pair of LPs. If (P) has an optimal
solution, then (D) has one, and their objective values equal.

(D)(P) optimal solution unbounded infeasible

optimal solution possible 1 impossible
2 impossible 3
unbounded impossible 4 impossible
5 possible 6
infeasible impossible 7 possible 8 possible 9

University
c of Waterloo 11 / 11