Scribd
Upload
2 views

10-1_duality

10-1_duality

Uploaded by

lisiyu820
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
2 views

10-1_duality

10-1_duality

Uploaded by

lisiyu820
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

10.

1 - Duality: definition
Optimisation problems turn out to be closely related to their dual problems. Studying
duality is useful to gain insight into the mathematical structure of the original problems
and sometimes offers more efficient ways to solve them. We will take a brief look at the
definition of Lagrange duality in the context of general optimisation problems before we
focus on its application to linear programming.

Lagrange Duality
Consider a general optimisation problem in standard form.

min f0 (x)
x
s.t. fi (x) ≤ 0 i = 1, . . . n (1)
hi (x) = 0 i = 1, . . . , m

with x ∈ D ⊂ Rp . Assume that its domain D is nonempty. We define the Lagrangian


associated with the problem to be
n
X m
X
L(x, λ, ν) = f0 (x) + λi fi (x) + νi hi (x) (2)
i=1 i=1

The scalars λi , νi are called Lagrange Multipliers and we call the associated vectors
λ ∈ Rn , and ν ∈ Rm , dual variables.

Lagrange dual function


We define the Lagrange dual function to be the greatest lower bound of the Lagrangian
n m
!
X X
g(λ, ν) = inf L(x, λ, ν) = inf f0 (x) + λi fi (x) + νi hi (x) . (3)
x∈D x∈D
i=1 i=1

When the Lagrangian is unbounded below, the dual function takes on a value of −∞.

Lower bound property


If λ ≥ 0 then
g(λ, ν) ≤ p∗ , (4)

1
where p∗ is the optimal value p∗ = f0 (x∗ ) of (1). To see this, suppose x̃ is a feasible
point for the problem (1), (i.e. that fi (x̃) ≤ 0 for all i = 1, . . . , n and hi (x̃) = 0 for all
i = 1, . . . , m) and that λ ≥ 0. Then
n
X m
X
L(x̃, λ, ν) = f0 (x̃) + λi fi (x̃) + νi hi (x̃) ≤ f0 (x̃).
i=1 i=1

It then follows that for any feasible point x̃, and λ ≥ 0 that

g(λ, ν) = inf L(x, λ, ν) ≤ L(x̃, λ, ν) ≤ f0 (x̃).


x∈D

In particular, if x̃ = x∗ then
g(λ, ν) ≤ p∗ .

The Lagrange dual problem


The Lagrange dual function therefore provides us with a lower bound on the optimal
objective function value f0 (x∗ ) for any λ ≥ 0, and ν. A natural question that follows
is then: “What is the best lower bound that can be obtained from the Lagrange dual
function?”. The associated optimisation problem is called the Lagrange dual problem:
max g(λ, ν)
λ,ν
(5)
s.t. λ ≥ 0.

We say that λ, ν are dual feasible if λ ≥ 0 and g(λ, ν) > −∞. Finally we refer to
(λ∗ , ν ∗ ) as dual optimal or optimal Lagrange multipliers if they are optimal for (5).

In this context, we call the original optimisation (1) the primal problem, or simply the
primal.

Duals of common LP forms


We’ll look at some examples of the Lagrange dual function for different Linear Program-
ming formulations.
Example 1 (Standard form LP). The standard form LP has the following form:

min cT x
x
s.t. Ax = b
x ≥ 0.

2
The Lagrangian is then given by

L(x, λ, ν) = cT x + ν T (Ax − b) − λT x
= −bT ν + (AT ν − λ + c)T x

and the Lagrange dual function is therefore given by

g(λ, ν) = −bT ν + inf (AT ν − λ + c)T x


x
(
T
−b ν if AT ν − λ + c = 0
=
−∞ otherwise

Note that in the second case above, since the domain of x is Rp (x ≥ 0 taken into
account as the constraint), the function value can be arbitrarily negative. Therefore
g(λ, ν) takes the value of −∞. The Lagrange dual problem (or simply the dual) is

max − bT ν
λ,ν

s.t. AT ν − λ + c = 0 (6)
λ≥0

which is simply another LP. This can also be expressed in inequality form, treating the λ
as slack variables:
max − bT ν
ν
(7)
s.t. − AT ν ≤ c.
Notice that ν is not restricted in sign. Hence, we can write ν in place of −ν and simplify
the problem. In other words, we can get rid of the negative signs in the inequality-form
LP.

Example 2 (Inequality form LP). Consider now an LP in inequality form:

min cT x
x
s.t. Ax ≤ b.

There are no equality constraints, so we only need λ Lagrange multipliers. The La-
grangian is given by

L(x, λ) = cT x + λT (Ax − b) = −bT λ + (AT λ + c)T x.

3
and the Lagrange dual is
(
− bT λ if AT λ = −c, λ ≥ 0
g(λ) =
−∞ otherwise

The dual problem is therefore


max − bT λ
λ
s.t. AT λ = −c (8)
λ≥0
which is, in fact, a standard form LP.

Note for both of these examples, that the number of primal constraints is equal to the
number of dual variables, and vice versa.

You might also like