Optimisation Theory

An Introduction to Optimisation
Dr Damien S. Eldridge
Australian National University
19 May 2017
D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 1 / 48

Sources Part 1
This material has been drawn from, or influenced by, a number of

sources. Some of these include those listed below.
Introduction-level references.
Bradley, T (2008), Essential mathematics for economics and business
(third edition), John Wiley and Sons, Great Britain: Chapter 7
(Section 4) (pp. 408-420).
Haeussler, EF Jr and RS Paul (1987), Introductory mathematical
analysis for business, economics, and the life and social sciences (fifth
edition), Prentice-Hall International Edition, Prentice-Hall, USA:
Chapter 17 (Section 8) (pp. 706-714).
Shannon, J (1995), Mathematics for business, economics and finance,
John Wiley and Sons, Brisbane: Chapter 10, Section 7 (pp. 489-501).
Sydsaeter, K, and P Hammond (2006), Essential mathematics for
economic analysis (second edition), Prentice Hall Financial Times
(Pearson Education), Gosport: Chapters 8, 13 and 14 (pp. 269-304
and 463-548).
Continued on the next slide.
Sources Part 2
Continued from the previous slide.
Intermediate-level references.
Chiang, AC and K Wainwright (2005), Fundamental methods of
mathematical economics (fourth edition), McGraw-Hill, Singapore:
Chapters 9 and 11-13 (pp. 220-254 and 291-442).
Dixit, AK (1990), Optimization in economic theory (second edition),
Oxford University Press, Great Britain: Chapters 1 to 9 and the
Appendix (pp. 1-144 and 181-185).
Leonard, D, and NV Long (1992), Optimal control theory and static
optimization in economics, Cambridge University Press, USA: Chapter
1 (pp. 1-86).
Silberberg, E (1990), The structure of economics: A mathematical
approach (second edition), McGraw-Hill, Singapore: Chapters 4, 6-14
and 17 (pp. 107-134, 156-490 and 573-612).
Simon, C, and L Blume (1994), Mathematics for economists, WW
Norton and Co, USA: Chapters 16-22, 23 (Section 8) and 30 (pp.
375-576, 626-629 and 822-844).
Continued on the next slide.
Sources Part 3
Continued from the previous slide.

Intermediate-level references (continued).
Starr, RM (1997), General equilibrium theory: An introduction,
Cambridge University Press, USA: (pp. 207-209).
Takayama, A (1993), Analytical methods in economics, The University
of Michigan Press, USA: Chapter 2-5, Appendix B and Appendix C
(pp. 75-321 and 605-647).
Advanced-level references.
Ausubel, LM and RJ Deneckere (1993), “A generalized theorem of the
maximum”, Economic Theory 3(1), January, pp. 99-107.
Corbae, D, MB Stinchcombe and J Zeman (2009), An introduction to
mathematical analysis for economic theory and econometrics, Princeton
University Press, USA: Chapters 5 and 6 (pp. 172-354).
Sundaram, RK (1996), A first course in optimization theory,
Cambridge University Press, USA: Chapters 2-10 (pp. 74-267).
Takayama, A (1985), Mathematical economics (second edition),
Cambridge University Press, USA: Chapters 1 and 2 (pp. 59-294).
Why do economists care about optimisation? Part 1
Economists try to explain social phenomena in terms of the behaviour of

an individual who is confronted with scarcity and the interaction of that
individual with other individuals who also face scarcity. This is perhaps
best captured by Malinvaude’s definition of economics:
”· · · economics is the science which studies how scarce
resources are employed for the satisfaction of the needs of men
living in society: on the one hand, it is interested in the essential
operations of production, distribution and consumption of goods,
and on the other hand, in the institutions and activities whose
object it is to facilitate these operations.” (Italics in original.)
(From page one of Malinvaude, E. (1972), Lectures on microeconomic

theory, Advanced Textbooks in Economics Volume 2, North Holland
Publishing Company, Scotland, translated by Mrs. A. Silvey.)

Scarcity:
This is the defining feature of economics.
It is this feature that distinguishes economics from other social sciences.
The behaviour of an individual who is faced with scarcity:
Often modelled using “constrained optimisation” techniques.
The interaction of individuals that face scarcity:
Economic equilibrium (eg competitive equilibrium and Nash
equilibrium).
When does a system of equations have at least one solution?
How do we find such a solution (if it exists)?
Use techniques from linear algebra and (for nonlinear cases) fixed point
theorems.

According to Ausubel and Deneckere (1993, p. 99):
“Almost every economic problem involves the study of an

agent’s optimal choice as a function of certain parameters or
state variables. For example, demand theory is concerned with
an agent’s optimal consumption as a function of prices and
income, while capital theory studies the optimal investment rule
as a function of the existing capital stock.”

Maximisation or Minimisation?
Optimisation problems come in two broad varieties: maximisation
problems and minimisation problems.
A maximisation problem involves finding a maximum for some function
f : S −→ R, where S ⊆ Rn , on some subset of its domain C ⊆ S.
A minimisation problem typically involves finding a minimum for some
function f : S −→ R, where S ⊆ Rn , on some subset of its domain
C ⊆ S.
Some terminology.
The function f is known as the objective function.
The set C is known as the constraint set (or the feasible set).
It turns out that the problem of finding a minimum value for f on the
set C is equivalent to the problem of finding a maximum value for
(−f ) on the set C .
Thus we can restrict our attention to maximisation problems without
loss of generality.
If we are faced with a minimisation problem, we can simply multiply the
objective function by (−1) and treat it like a maximisation problem.
Unconstrained or Constrained?
Consider an optimisation problem that involves finding a maximum

for some function f : S −→ R, where S ⊆ Rn , on some subset of its
domain C ⊆ S.
If C = S, then the problem is an unconstrained optimisation problem.
If C ⊂ S (that is, if C ⊆ S and C 6= S), then the problem is a
constrained optimisation problem.
In this course, we will only look at unconstrained problems and
equality-constrained problems. But please note that
inequality-constraints are also possible. These are covered in more
advanced courses.

The existence and properties of a solution
As we shall see, there are some optimisation problems that have no

solutions.
It would be nice to know that the problem we are trying to solve has
a solution before we attempt to solve it.
In cases where f is a function, a set of conditions that are sufficient
to guarantee the existence of a solution to an optimisation problem is
provided by Weierstrass’ Theorem of the Maximum.
A set of conditions that are sufficient to guarantee that that the
solution set and the set of optimal choice variables in an optimisation
problem have some desirable properties is provided by Berge’s
Theorem of the maximum.
These two theorems are beyond the scope of this course. However, it
is comforting to know that they exist!

The potential non-existence of a solution
We will illustrate the potential non-existence of a solution by means
of some examples.
Consider the following optimisation problems.
Find a maximum for f (x ) = x when x ∈ (−∞, ∞).
No solution exists because f (x ) → ∞ as x → ∞.
Find a minimum for f (x ) = x when x ∈ (−∞, ∞).
No solution exists because f (x ) → −∞ as x → −∞.
Find a maximum for f (x ) = x when x ∈ (a, b ) where
−∞ < a < b < ∞.
No solution exists because of the constraint set is open.
f (x ) → f (b ) as x → b, but never quite gets there.
What is the largest real number that is strictly less than b?
Find a minimum for f (x ) = x when x ∈ (a, b ) where
−∞ < a < b < ∞.
No solution exists because of the constraint set is open.
f (x ) → f (a) as x → a, but never quite gets there.
What is the largest real number that is strictly greater than a?
Illustrate these examples on the white-board.
Solutions and Control Variables Part 1
The solution to a constrained maximisation problem is the maximum

value of the objective function f (x1 , x2 , · · · , xn ) over the constraint
set C
It is denoted by the maximum value function
f ∗ (·) ≡ max {f (x ) : x ∈ C } .
The optimal choices of the xi variables are known as ”arg max”s:

 ∗ 
 x1 (·) 
 ∗
 
∗

 2x (·) 
{ (·)}
x =

 .. 

  .  
xn∗ (·)

 

= arg max {f (x ) : x ∈ C } .

We can find the maximum value function from the ”arg max”s as
follows:
f ∗ (·) = f (x ∗ )
where
x ∗ ∈ arg max {f (x ) : x ∈ C } .
If the ”arg max” is unique, then this becomes
f ∗ (·) = f (x ∗ )
where
x ∗ = arg max {f (x ) : x ∈ C } .

The solution to a constrained minimisation problem is the minimum

value of the objective function f (x1 , x2 , · · · , xn ) over the constraint
set C .
It is denoted by the minimum value function
f ∗ (·) ≡ min {f (x ) : x ∈ C } .
The optimal choices of the xi variables are known as ”arg min”s:

 ∗ 
 x 1 (·) 
 ∗
 
∗

 2x (·) 
{ (·)}
x =

 .. 

  .  
xn∗ (·)

 

= arg min {f (x ) : x ∈ C } .

We can find the minimum value function from the ”arg min”s as
follows:
f ∗ (·) = f (x ∗ )
where
x ∗ ∈ arg min {f (x ) : x ∈ C } .
If the ”arg min” is unique, then this becomes
f ∗ (·) = f (x ∗ )
where
x ∗ = arg min {f (x ) : x ∈ C } .

Global Optima and Local Optima Part 1
Let f : S −→ R and C ⊆ S, where S ⊆ Rn .

Global and local maxima.
f (x ∗ ) is a global maximum of f on C if x ∗ ∈ C and f (x ∗ ) > f (x ) for
all x ∈ C .
f (x ∗ ) is a local maximum of f on C if x ∗ ∈ C and f (x ∗ ) > f (x ) for
all x ∈ C that are “sufficiently close to” x ∗ .
Global and local minima.
f (x ∗ ) is a global minimum of f on C if x ∗ ∈ C and f (x ∗ ) 6 f (x ) for
all x ∈ C .
f (x ∗ ) is a local minimum of f on C if x ∗ ∈ C and f (x ∗ ) 6 f (x ) for all
x ∈ C that are “sufficiently close to” x ∗ .

Global Optima and Local Optima Part 2
Note that every global maximum is a local maximum, but not all
local maxima will be a global maximum.
Illustrate this on the white-board.
Note that every global minimum is a local minimum, but not all local
minima will be a global minimum.
Illustrate this on the white-board.

Unconstrained Optimisation
Let f : S −→ R be a function that is at least twice continuously
differentiable and C = S, where S ⊆ Rn is an open set.
Since C = S, this an unconstrained optimisation proiblem.
The point x ∗ ∈ S will yield a local maximum of f if both of the
following sets of conditions are satisfied.
(First-Order Conditions) (FOC): The gradient vector Dx f equals the
null vector at the point x ∗ . (In other words, Dx f (x ∗ ) = 0.
∂f (x ∗ )
This requires that ∂xi = 0 for all i ∈ {1, 2, · · · , n}.
2 f (x ∗ ) is
(Second-Order Conditions) (SOC): The Hessian matrix Dxx T
negative definite at the point x ∗
The point x ∗ ∈ S will yield a local minimum of f if both of the

following sets of conditions are satisfied.
(First-Order Conditions) (FOC): The gradient vector Dx f equals the
null vector at the point x ∗ . (In other words, Dx f (x ∗ ) = 0.
∂f (x ∗ )
This requires that ∂xi = 0 for all i ∈ {1, 2, · · · , n}.
2 f (x ∗ ) is
(Second-Order Conditions) (SOC): The Hessian matrix Dxx T
positive definite at the point x ∗

Univariate Unconstrained Optimisation Examples
Work through the following examples on the white-board.

Minimise f (x ) = (x − 1)2 on R. (A minimum of f = 0 at x = 1.)
x − 1) on R. (A maximum of f = 0 at x = 1.)
Maximise f (x ) = −( 2
Maximise f (x ) = 13 x 3 − x + 1 on R. (A “ minimum” turning point

at x = 1 and a “maximum” turning point at x = −1. A local
maximum of f = 35 at x = −1.)
Minimise f (x ) = 3 x 3 − x + 1 on R. (A “ minimum” turning point
1
at x = 1 and a “maximum” turning point at x = −1. A local

minimum of f = 31 at x = 1.)

A Multivariate Unconstrained Optimisation Example
Consider the function f : R2 −→ R defined by f (x, y ) = −x 2 − y 2 .
The first-order conditions for either a maximum or a minimum are:
fx (x, y ) = −2x = 0 =⇒ x ∗ = 0; and
fy (x, y ) = −2y = 0 =⇒ y ∗ = 0.
Thus (x ∗ , y ∗ ) = (0, 0) is the unique critical point for f (x, y ).
The Hessian matrix for f (x, y ) is

fxx (x, y ) fxy (x, y ) −2 0
H= = .
fyx (x, y ) fyy (x, y ) 0 −2
Note that:
H is a symmetric matrix,
det(H1 ) = −2 < 0, and
det(H2 ) = det(H ) = 4 > 0.
Thus H is negative definite for all (x, y ) ∈ R2 .
This means that (x ∗ , y ∗ ) = (0, 0) yields a global maximum of f (x, y ).
The maximum value of f (x, y ) is f ∗ = f (x ∗ , y ∗ ) = f (0, 0) = 0.
Univariate Profit Maximisation by a Price-Taking Firm
Part 1
Suppose that a single-product price-taking firm faces an output price
of P per unit and has a (total) cost function that is given by C (Q ).
We will assume that the cost function is at least twice continuously
differentiable.
The firm’s profits are given by the function Π(Q ) = PQ − C (Q ).
Suppose that this firm wants to choose Q to maximise its profits.
We will ignore the shut-down condition as a matter of convenience.
The first-order condition for profit maximisation is
dΠ dC
=P− = 0.
dQ dQ
dC
This can be rearranged to obtain P = MC (Q ), where MC (Q ) = dQ .
This first-order condition implicitly defines one or more critical values
of Q for the function Π(Q ).
Univariate Profit Maximisation by a Price-Taking Firm
Part 2
Suppose that Q ∗ is a critical value of Q for the function Π(Q ).

The second-order condition for a local maximum requires that
d 2Π d 2C dMC
2
= − 2
=− < 0,
dQ dQ dQ
when it is evaluated at the point Q ∗ .
Note that this requires that dMC
dQ > 0 when it is evaluated at the point
Q ∗.
In other words, we need marginal cost to be an increasing function of
the output quantity in a non-empty neighbourhood around the point
Q ∗.

Bivariate Profit Maximisation by a Price-Taking Firm Part
1
Suppose that a two-product price-taking firm faces output prices of

P1 per unit for one of its products and and P1 per unit for its other
product.
Suppose also that this firm has a (total) cost function that is given by
C (Q1 , Q2 ), where Q1 is its output of good one and Q2 is its output
of good two.
We will assume that the cost function is at least twice continuously
differentiable in all of its arguments.
The firm’s profits are given by the function
Π(Q1 , Q2 ) = P1 Q1 + P2 Q2 − C (Q1 , Q2 ).
Suppose that this firm wants to choose Q1 and Q2 to maximise its
profits.
We will ignore the shut-down condition as a matter of convenience.

2
The first-order conditions for profit maximisation by the firm are
∂Π ∂C
= P1 − =0
∂Q1 ∂Q1
and
∂Π ∂C
= P2 − = 0.
∂Q2 ∂Q2
These two conditions can be rearranged to obtain
P1 = MC1 (Q1 , Q2 )
and
P2 = MC2 (Q1 , Q2 ),
where MCi (Q1 , Q2 ) = ∂C
∂Qi for i ∈ {1, 2}.
3
The two first-order conditions for this bivariate profit maximisation

problem jointly define one or more critical points for the function
Π(Q1 , Q2 ).
Let Q ∗ = (Q1∗ , Q2∗ ) be one such point.
The second-order condition for Q ∗ to yield a local maximum of the
function Π(Q1 , Q2 ) requires that the Hessian matrix for Π(Q1 , Q2 )
be negative definite when it is evaluated at the point Q ∗ = (Q1∗ , Q2∗ ).

4
The Hessian matrix for the function Π(Q1 , Q2 ) is
H = D 2 Π(Q1 , Q2 )
 
2 ∂ Π 2 ∂ Π
∂Q12 ∂Q2 ∂Q1
= ∂2 Π ∂2 Π

∂Q1 ∂Q2 ∂Q22
 2 2

∂ C C
− ∂Q 2 − ∂Q∂2 ∂Q 1
= C2
1
∂ C 2

− ∂Q∂1 ∂Q2
− ∂Q 2
2
!
− ∂MC 1
− ∂MC 1
= ∂Q1 ∂Q2 .
− ∂Q1
∂MC2
− ∂Q2
∂MC2

5
A sufficient condition for the critical point Q ∗ = (Q1∗ , Q2∗ ) to yield a

local maximum of the function Π(Q1 , Q2 ) is that the Hessian matrix
(H) be negative definite at the point Q ∗ = (Q1∗ , Q2∗ )
Assuming that the function Π(Q1 , Q2 ) is at least twice continuously
differentiable in a non-empty neighbourhood around the point
Q ∗ = (Q1∗ , Q2∗ ) (so that Young’s theorem applies and the Hessian
matrix is symmetric at the critical point being considered), this
requires that:
1 det(H ) = −
∂MC1 ∗ ∗ ∗
1 ∂Q1 < 0 at Q = (Q1 , Q2 ), and

2 det(H2 ) = det(H ) = ∂MC
∂Q1
1 ∂MC2
∂Q2 − ∂MC∂Q2
1 ∂MC2
∂Q1 =
2
∂MC1
∂Q1
∂MC2
∂Q2 − ∂MC
∂Q2
1
> 0 at Q ∗ = (Q1∗ , Q2∗ ).

6
This in turn requires that:

∂MC1 ∗ ∗ ∗
∂Q1 > 0 at Q = (Q1 , Q2 ),
1
∂MC2 ∗ ∗ ∗
∂Q2 > 0 at Q = (Q1 , Q2 ), and
2
2
3
∂MC1
∂Q1
∂MC2
∂Q2 > ∂MC1
∂Q2 at Q ∗ = (Q1∗ , Q2∗ ).

Equality Constrained Optimisation
Let f : S −→ R be a function that is at least twice continuously

differentiable and C ⊆ S, where S ⊆ Rn is an open set.
Suppose that the constraint set C consists only of equality
restrictions. In other words, suppose that
C = x ∈ Rn : g i (x ) = b i for all i ∈ {1, 2, · · · , m } ,

where b i ∈ R for i ∈ {1, 2, · · · , m } are all constant terms.

Consider the problem of maximising f on C .
We can resrict our attention to a constrained maximisation problem
without loss of generality.
The reason for this is that minimising f on C is equivalent to
maximising (−f ) on C .
Note that a necessary condition for their to be a solution to a
constrained optimisation problem is that the constraint set be
non-empty.
The Lagrangian Function
Let S ⊆ Rn be an open set, f : S −→ R be a function that is at least

twice continuously differentiable, and
C = x ∈ Rn : g i (x ) = b i for all i ∈ {1, 2, · · · , m } ,

be a constraint set.
Consider the problem of maximising f on C .
The Lagrangian function for this constrained maximisation problem is
m
L (x, λ) = f (x ) + ∑ λi bi − g i (x ) .

i =1
The λi variables are known as Lagrange multipliers.

Equality Constrained Optimisation Solutions
Theorem: x ∗ ∈ C yields a local maximum of f if the following three

sets of conditions are satisfied.
(FOC Part 1): Dx L(x ∗ , λ∗ ) = 0.
(FOC Part 2): Dλ L(x ∗ , λ∗ ) = 0.
An appropriate rank condition and appropriate second-order conditions
are satisfied. (These are beyond the scope of this course.)

The First-Order Conditions for Equality-Constrained
Maximisation Problems
Recall that the Lagrangian function for the constrained maximisation

problem under consideration is
m
L (x, λ) = f (x ) + ∑ λi bi − g i (x ) .

i =1
The first-order conditions for this problem are:

∂L m
∂xk + ∑i =1 − λi ∂xk = 0 for each k ∈ {1, 2, · · ·
∂g
∂xk = ∂f
, n }, and
∂L
∂λj = b j − g j (x ) = 0 for each j ∈ {1, 2, · · · , m }.

The Second-Order Conditions for Equality-Constrained
Maximisation Problems
It is possible to express the second-order conditions for this problem

in terms of something known as a bordered Hessian matrix. However,
this is beyond the scope of this course.
We will instead look at converting equality-constrained optimisation
problems into unconstrained optimisation problems in fewer variables
by using the constraints to express some variables in terms of others.
This will allow us to use the second-order conditions for uncontrained
optimisation problems.

Interpreting Lagrange Multipliers
The optimal value of the Lagrange multiplier for the ith constraint
(that is, the value of that Lagrange multiplier when all of the
first-order conditions are satisfied) tells us the degree of sensitivity of
the optimal value of the objective function over the constraint set,
f ∗ (x∗ , b ), to a change i
in the ith constraint value b , where
1
b = b ,b ,··· ,b2 m is the vector of constraint values.
It is, in effect, a“shadow price”.
The optimal value of the Lagrange multiplier in a standard utility
maximisation problem can be interpreted as the marginal utility of
income.
The optimal value of the Lagrange multiplier in a cost minimisation
problem can be interpreted as marginal cost.

Some Economic Applications
Budget-contrained utility maximisations problems.

Utlity-constrained expenditure minimisation problems.
Output-constrained cost minimisation problems.
Technology-constrained profit maximisation problems.
Pareto efficiency problems.

Equality Constrained Optimisation Examples
Work through the following examples on the board or from other

notes.
Maximise f (x, y ) = 3x − y + 6 subject to the constraint that
x 2 + y 2 = 4, where (x, y ) ∈ R2 .
Solve some UMP examples.
Solve a cost-minimisation problem for a particular two-input version of
a Cobb-Douglas production function and show that the optimal value
of the Lagrange multiplier on the “minimum output” constraint is
equal to marginal cost.

Converting Equality Constrained Optimisation Problems
into Unconstrained Optimisation Problems Part 1
Consider an equality constrained optimisation problem of the form
Max f (x1 , x2 , · · · , xn ) subject to
g 1 (x1 , x2 , · · · , xn ) = b1 ,
g 2 (x1 , x2 , · · · , xn ) = b2 ,
..
.
g m (x1 , x2 , · · · , xn ) = bm .
The constraint set for this problem is C =
(x1 , x2 , · · · , xn ) : g j (x1 , x2 , · · · , xn ) = bj for all j ∈ {1, 2, · · · , m} .

Assume that there are fewer constraints than there are choice
variables (m < n) and that the constraint set is nonempty (C 6= ∅).

Since m < n and C 6= ∅, we know that if the constraint equations

are all linear and the domain of each choice variable is an infinite
subset of R, then there will be an infinite number of solutions for the
system of equations given by
g 1 (x1 , x2 , · · · , xn ) = b1 ,
g 2 (x1 , x2 , · · · , xn ) = b2 ,
..
.
g m (x1 , x2 , · · · , xn ) = bm .
These solutions would be parametric. They would pin down the value
of m of the choice variables in terms of the other (n − m ) choice
variables and the constraint values. However, each of the remaining
(n − m) choice variables would be free to vary over their entire
domain.
Suppose that this is the case even when the constraint equations are
nonlinear. (This may impose some restrictions on the constraint
equations. Think about the implicit function theorem and the inverse
function theorem, for example.)
In this case, we can use the system of constraint equations to express
m of the choice variables as implicit functions of the remaining
(n − m) choice variables and the m constraint values.
These implicit functions might take the form
xn−m+1 = hn−m+1 (x1 , x2 , · · · , xn−m , b1 , b2 , · · · , bm ),

xn−m+2 = hn−m+2 (x1 , x2 , · · · , xn−m , b1 , b2 , · · · , bm ),
..
.
xn = hn (x1 , x2 , · · · , xn−m , b1 , b2 , · · · , bm ).

Having used the system of constraint equations to express m of the

choice variables as implicit functions of the remaining (n − m ) choice
variables and the m constraint values, we can now substitute these
implicit functions back into the objective function to obtain the
following “reduced-form” unconstrained optimisation problem:
Max f (x1 , x2 , · · · , xn−m , hn−m+1 (·), hn−m+2 (·), · · · , hn (·)).
Note that there are only (n − m ) choice variables in this

“reduced-form” unconstrained optimisation problem.
But each of these variables potentially enters the objective function
multiple times; once directly and up to m times indirectly. The
indirect entries come through the implicit functions that solve the set
of constraint equations (in other words, the h functions).

The technique for solving unconstrained optimisation problems that

we have already covered can now be applied to this ‘reduced-form”
unconstrained optimisation problem in order to solve the oringinal
equality-constrained optimisation problem.
Once optimal values for the (n − m ) choice variables in this
“reduced-form” unconstrained optimisation problem have been
obtained, the optimal values of remaining m choice variables from the
original equality-constrained optimisation problem can be obtained by
substituting them into the h functions. This will give us the optimal
choice values for all n of the choice variables.
The solution to the equality-constrained optimisation problem can
then be found by substituting the optimal values for all n of the
choice variables into the objective function.

Conversion Example Part 1
Consider the following UMP:
Max U (x1 , x2 , x3 ) = αln(x1 ) + βln(x2 ) + (1 − α − β)ln(x3 )
subject to p1 x1 + p2 x2 + p3 x3 = y ,
where 0 < α < 1, 0 < β < 1, and α + β < 1.
We will assume throughout that x1 > 0, x2 > 0, and x3 > 0.
We can use the budget constraint to express x3 as a function of x1
and x2 as follows:
y p1 p2
x3 = − x1 − x2 .
p3 p3 p3
Upon substituting this into the utility function, we obtain the
following unconstrained optimisation problem:
Max Û (x1 , x2 ) =
y p1 p2
αln(x1 ) + βln(x2 ) + (1 − α − β)ln( − x1 − x2 ).
p3 p3 p3
The first-order conditions for this problem are
∂Û α (1 − α − β) pp13
= − y = 0,
∂x1 x1 ( p3 − pp13 x1 − pp23 x2 )
and
∂Û β (1 − α − β) pp23
= − y = 0.
∂x2 x2 ( p3 − pp13 x1 − pp23 x2 )
These can be rearranged and simplified to obtain
(1 − α − β)p1 x1 = α(y − p1 x1 − p2 x2 ),
and
(1 − α − β)p2 x2 = β(y − p1 x1 − p2 x2 ).

These can be further rearranged and simplified to obtain the following
system of two equations in two unknowns:
(1 − β)p1 x1 + αp2 x2 = αy ,
and
βp1 x1 + (1 − α)p2 x2 = βy .
It is probably easiest to solve this system of equations by using
expenditure on each commodity as the variables. Upon doing this, we
obtain p1 x1 = αy and p2 x2 = βy , so that
αy βy
x1 = and x2 = .
p1 p2
This means that
y p1 αy p2 βy (1 − α − β )y
x3 = − − = .
p3 p3 p1 p3 p2 p3
Thus our candidate solution is
V (p1 , p2 , p3 , y ) = U ∗
αy βy (1 − α − β)y

=U , ,
p1 p2 p3
(1 − α − β )y

αy βy
= αln + βln + (1 − α − β)ln
p1 p2 p3
!
( 1 − α − β ) y (1− α − β )
α β
αy βy
= ln .
p1 p2 p3
What about second-order conditions?
We need to examine the definiteness (or otherwise) of the Hessian
matrix for the utility function from the “reduced-form” unconstrained
optimisation problem.
Recall that
α (1 − α − β) pp31
Û1 = − y = 0,
x1 ( p3 − pp13 x1 − pp23 x2 )
and
β (1 − α − β) pp32
Û2 = − y = 0.
x2 ( p3 − pp13 x1 − pp23 x2 )
Thus we have
2
p1
α (1 − α − β ) p3
Û11 =− 2− y p1 p2 2
< 0 for all x1 and x2 ,
x1 ( p3 − p3 x1 − p3 x2 )

p1 p2
(1 − α − β ) p32
Û12 = Û21 = − , and
( py3 − p1
p3 x1 − p2
p3 x 2 )
2
Continued on next page.

Continued from previous page

2
p2
β (1 − α − β ) p3
Û22 =− 2− y p1 p2 2
.
x2 ( p3 − p3 x 1 − p3 x2 )
Recall that
Û11 Û12
H= .
Û21 Û22
Note that
det(H1 ) = det(Û11 ) = Û11
2
p1
α ( 1 − α − β ) p3
=− 2− y p1 p2 < 0 for all x1 and x2 , and
x1 ( p3 − p3 x1 − p3 x2 )2
Continued on next page.

Continued from previous page
det(H2 ) = det(H ) = Û11 Û22 − Û12 Û21 .
After some algebra and simplification, this becomes:

2 2
p2 p1
αβ (1 − α − β ) p3 αx12 (1 − α − β ) p3 βx22
det(H2 ) = 2 2 + +
x1 x2 ( py3 − p1 p2
p3 x1 − p3 x2 )
2 ( py3 − p1 p2
p3 x1 − p3 x2 )
2
> 0 for all x1 and x2 .

This means that the hessian matrix is negative definite for all
combinations of (x1 > 0 and x2 > 0). Thus we can conclude that our
candidate solution is a global maximum for the optimisation problem
under consideration.

Optimisation Theory

Uploaded by

Copyright:

Available Formats

Optimisation Theory

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Optimisation Theory

Uploaded by

Copyright:

Available Formats

An Introduction to Optimisation

Australian National University

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 1 / 48

This material has been drawn from, or influenced by, a number of

Continued from the previous slide.

Economists try to explain social phenomena in terms of the behaviour of

(From page one of Malinvaude, E. (1972), Lectures on microeconomic

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 5 / 48

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 6 / 48

According to Ausubel and Deneckere (1993, p. 99):

“Almost every economic problem involves the study of an

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 7 / 48

Consider an optimisation problem that involves finding a maximum

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 9 / 48

As we shall see, there are some optimisation problems that have no

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 10 / 48

The solution to a constrained maximisation problem is the maximum

The optimal choices of the xi variables are known as ”arg max”s:

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 12 / 48

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 13 / 48

The solution to a constrained minimisation problem is the minimum

The optimal choices of the xi variables are known as ”arg min”s:

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 14 / 48

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 15 / 48

Let f : S −→ R and C ⊆ S, where S ⊆ Rn .

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 16 / 48

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 17 / 48

The point x ∗ ∈ S will yield a local minimum of f if both of the

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 18 / 48

Work through the following examples on the white-board.

Maximise f (x ) = 13 x 3 − x + 1 on R. (A “ minimum” turning point

at x = 1 and a “maximum” turning point at x = −1. A local

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 19 / 48

Suppose that Q ∗ is a critical value of Q for the function Π(Q ).

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 22 / 48

Suppose that a two-product price-taking firm faces output prices of

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 23 / 48

The first-order conditions for profit maximisation by the firm are

The two first-order conditions for this bivariate profit maximisation

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 25 / 48

The Hessian matrix for the function Π(Q1 , Q2 ) is

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 26 / 48

A sufficient condition for the critical point Q ∗ = (Q1∗ , Q2∗ ) to yield a

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 27 / 48

This in turn requires that:

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 28 / 48

Let f : S −→ R be a function that is at least twice continuously

C = x ∈ Rn : g i (x ) = b i for all i ∈ {1, 2, · · · , m } ,

where b i ∈ R for i ∈ {1, 2, · · · , m } are all constant terms.

Let S ⊆ Rn be an open set, f : S −→ R be a function that is at least

C = x ∈ Rn : g i (x ) = b i for all i ∈ {1, 2, · · · , m } ,

The λi variables are known as Lagrange multipliers.

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 30 / 48

Theorem: x ∗ ∈ C yields a local maximum of f if the following three

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 31 / 48

Recall that the Lagrangian function for the constrained maximisation

The first-order conditions for this problem are:

D. S. Eldridge (ANU) An Introduction to Optimisation 19 May 2017 32 / 48