Chapter 1

Further topics in optimiza-

tion

Maximum Value Functions and the

Envelope Theorem1
A maximum (or minimum) value function is an objective function
where the choice variables have been assigned their optimal values.
These optimal values of the choice variables are, in turn, functions
of the exogenous variables and parameters of the problem. Once the
optimal values of the choice variables have been substituted into the
original objective function, the function indirectly becomes a function
of the parameters (through the parameters’ influence on the optmal
values of the choice variables). Thus the maximum value function is
also referred to as the indirect objective function.
What is the significance of the indirect objective function? Con-
sider that in any optimization problem the direct objective function is
maximized (or minimized) for a given set of parameters. The indirect
objective function gives all the maximum values of the objective func-
1
This section of the chapter presents an overview of the envelope theorem for the
purpose of introducing the concept to the student. For a richer treatment of this
topic is found in chapter 7 of The Structure of Economics: a Mathematical Analysis
(3rd Ed.) by Eugene Silberberg and Wing Suen (McGraw-Hill, 2001) from which
parts of this section are based.

1
2 Further topics in optimization

tion as these prameters vary. Hence the indirect objective function is

an ”envelope” of the set of optimized objective functions generated by
varying the parameters of the model. For most students of economics
the first illustration of this notion of an ”envelope” arises in the com-
parison of short-run and long-run cost curves. Students are typically
taught that the long-run average cost curve is an envelope of all the
short-run average cost curves (what parameter is varying along the en-
velope in this case?). A formal derivation of this concept is one of the
exercises we will be considering in the following sections.
To illustrate, consider the following maximization problem with two
choice variables x and y, and one parameter, α:
Maximize
U = f (x, y, α) (1.1)
The first order necessary condition are
fx (x, y, α) = fy (x, y, α) = 0 (1.2)
if second-order conditions are met, these two equations implicitly define
the solutions
x = x∗ (α) y = x∗ (α) (1.3)
If we subtitute these solutions into the objective function, we obtain a
new function
V (α) = f (x∗ (α), y ∗ (α), α) (1.4)
where this function is the value of f when the values of x and y
are those that maximize f (x, y, α). Therefore, V (α) is the maximum
value function (or indirect objective function). If we differentiate V
with respect to α
∂V ∂x∗ ∂y ∗
= fx + fy + fα (1.5)
∂α ∂α ∂α
However, from the first order conditions we know fx = fy = 0.
Therefore, the first two terms disappear and the result becomes
∂V
= fα (1.6)
∂α
This result says that, at the optimum, as α varies, with x∗ and y ∗
allowed to adjust optimally gives the same result as if x∗ and y ∗ were
held constant! Note that α enters maximum value function (equation
1.4) in three places: one direct and two indirect (through x∗ and y ∗ ).
Equations 1.5 and 1.6 show that, at the optimimum, only the direct
effect of α on the objective function matters. This is the essence of the
envelope theorem. The envelope theorem says only the direct effects of
a change in an exogenous variable need be considered, even though the
exogenous variable may enter the maximum value function indirectly
as part of the solution to the endogenous choice variables.
Further topics in optimization 3

The Profit Function

Let’s apply the above approach to an economic application, namely the
profit function of a competitive firm. Consider the case where a firm
uses two inputs: capital, K, and labour, L. The profit function is

π = pf (K, L) − wL − rK (1.7)

where p is the output price and w and r are the wage rate and rental
rate respectively.
The first order conditions are
π L = fL (K, L) − w = 0
(1.8)
π K = fK (K, L) − r = 0
which respectively define the factor demand equations
L = L∗ (w, r, p)
(1.9)
K = K ∗ (w, r, p)
substituting the solutions K ∗ and L∗ into the objective function
gives us
π∗ (w, r, p) = pf (K ∗ , L∗ ) − wL∗ − rK ∗ (1.10)
π∗ (w, r, p) is the profit function (or indirect objective function). The
profit function gives the maximum profit as a function of the exogenous
variables w, r, and p.
Now consider the effect of a change in w on the firm’s profits. If we
differentiate the original profit function (equation 1.7) with respect to
w, holding all other variables constant and we get
∂π
= −L (1.11)
∂w
However, this result does not take into account the profit maximiz-
ing firms ability to make a substitution of capital for labour and adjust
the level of output in accordance with profit maximizing behavior.
Since π ∗ (w, r, p) is the maximum value of profits for any values of w,
r, and p, changes in π ∗ from a change in w takes all captial for labour
subsitutions into account. To evaluate a change in the maximum profit
function from a change in w, we differentiate π ∗ (w, r, p) with respect
to w yielding
∂π∗ ∂L∗ ∂K ∗
= [pfL − w] + [pfK − r] − L∗ (1.12)
∂w ∂w ∂w
From the first order conditions, the two bracketed terms are equal
to zero. Therefore, the resulting equation becomes
∂π ∗
= −L∗ (w, r, p) (1.13)
∂w
4 Further topics in optimization

This result says that, at the the profit maximizing position, a change
in profits with respect to a change in the wage is the same whether or
not the factors are held constant or allowed to vary as the factor price
changes. In this case the derivative of the profit function with respect
to w is the negative of the factor demand function L∗ (w, r, p). Follow-
ing the above procedure, we can also show the additional comparative
statics results
∂π ∗ (w, r, p)
= −K ∗ (r, w, p) (1.14)
∂r
and
∂π ∗ (w, r, p)
= f (K ∗ , L∗ ) = q∗ (1.15)
∂p
The simple comparative static results derived from the profit func-
tion is known as ”Hotelling’s Lemma”. Hotelling’s Lemma is simply
an application of the envelope theorem.

Reciprocity Conditions
Consider again our two variable maximization problem
Maximize U = f (x, y, α)
where x and y are the choice variable, and α is a parameter. The first
order equations are fx = fy = 0. which imply the functions x = x∗ (α)
and y = y ∗ (α).
We are interested in the comparative statics regarding the directions
of change in x∗ (α) and y ∗ (α)as α changes and the implications to the
the value function. The maximum value function is
V (α) = f (x∗ (α), y ∗ (α), α) (1.16)
By defination, V (α) gives the maximum value of f for a given α.
Now consider the new function, sometimes called the primal-dual
objective function, the difference between the actual and maximum
value for a given α,
Z = Z(x, y, α) = f (x, y, α) − V (α) (1.17)
This new function, Z, has a maximum of zero when x = x∗ , y = y ∗
and for any x 6= x∗ , y 6= y ∗ we find that f ≤ V . In this framework
Z(x, y, α) can be considered a function of three independent variables,
x, y, and α. The maximum of Z(x, y, α) = f (x, y, α) − V (α) can be
described by the first and second order conditions.
The first order conditions are:
Zx (x, y, α) = fx = 0
(1.18)
Zy (x, y, α) = fy = 0
Further topics in optimization 5

and
Zα (x, y, α) = fα − Vα = 0 (1.19)
We can see that the first-order conditions of our new function Z are
nothing more that are the original maximum conditions for f (x, y, α)
(equations 1.18). and the envelope theorem (equation 1.19) . These
first order conditions hold whenever x = x∗ (α) and y = y ∗ (α) The
sufficient second order conditions are
¯ ¯
¯ fxx fxy fxα ¯
¯ ¯
¯
H = ¯ fyx fyy fyα ¯ (1.20)
¯
¯ fαx fαy fαα − Vαα ¯

where the Hessian Matrix is negative definite, or

2
fxx < 0, fxx fyy − fxy > 0, H < 0 (1.21)

In addition, if the the second-order conditions are met, then fαα −

Vαα < 0 is also implied. This inequality is an important result which
plays an essential role in many comparitive static exercises. We know
already that
Vα (α) = fα (x∗ (α), y ∗ (α), α) (1.22)
Differentiating both sides with respect to α yields
∂x∗ ∂y ∗
Vαα = fαx − fαy + fαα (1.23)
∂α ∂α
From the sufficient second order conditions and using Young’s the-
orem
∂x∗ ∂y ∗
Vαα − fαα = fxα + fyα >0 (1.24)
∂α ∂α
Suppose that α enters only in the x first order condition such that
fyα = 0 Then equation 1.24 reduces to

∂x∗
fxα >0 (1.25)
∂α
∗
which implies that fxα and ∂x
∂α
will have the same sign2 .
For example, in the profit maximization model:

π = pf (K, L) − wL − rK (1.26)

Where the first order conditions are

π L = pfL − w = 0
(1.27)
π K = pfK − r = 0
2
This analysis is easily generalized to the n-variable case.
6 Further topics in optimization

The exogenous variable w enters only the first order equation pfL −
w = 0; it enters with a negative sign
∂π L
= −1 (1.28)
∂w
Therefore we can conclude that ∂L∗ /∂w will also be negative. Fur-
ther, if we combine of the envelope theorem with with Young’s theorem,
∗ ∗
we can show the reciprocity condition ∂L ∂r
= ∂K
∂w
. From the indirect
∗
profit function π (w, r, p) Hotelling’s Lemma gave us
π ∗w = ∂π∗
∂w
= −L∗ (w, r, p)
∂π∗ (1.29)
π r = ∂r = −K ∗ (w, r, p)
∗

differentiating again and applying Young’ theorem

∂L∗ ∂K ∗
π ∗wr = − =− = π ∗rw (1.30)
∂r ∂w
or
∂L∗ ∂K ∗
= (1.31)
∂r ∂w

The Envelope Theorem and Constrained Optimiza-

tion
Now let us turn our attention to the case of constrained optimization.
Again we will have an objective function (U), two choice variables, (x
and y) and one prarameter (α) except now we introduce the following
constraint:
g(x, y; α) = 0
The derivation of the envelope theorem for the models with one
constraint is as follows:
The problem then becomes
Maximize
U = f (x, y; α) (1.32)
subject to
g(x, y; α) = 0 (1.33)
The Lagrangian for this problem is

Z = f (x, y; α) + λg(x, y; α) (1.34)

The first order conditions are

Zx = fx + λgx = 0
Zy = fy + λgy = 0 (1.35)
Zλ = g(x, y; α) = 0
Further topics in optimization 7

Solving this system of equations gives us

x = x∗ (α) y = y ∗ (α) λ = λ∗ (α) (1.36)

Substituting the solutions into the objective function, we get

U ∗ = f (x∗ (α), y ∗ (α), α) = V (α) (1.37)

where V (α) is the indirect objective function, or maximum value

function. This is the maximum value of y for any α and xi ’s that satisfy
the constraint.

How does V (α) change as α changes? First, we differentiate V with

respect to α
∂V ∂x∗ ∂y ∗
= fx + fy + fα (1.38)
∂α ∂α ∂α
In this case,equation 1.38 will not simplify to ∂V
∂α
= fα since fx 6= 0
and fy 6= 0. However, if we substitute the solutions to x and y into the
constraint (producing an identity)

g(x∗ (α), y ∗ (α), α) ≡ 0 (1.39)

and differentiating with respect to α yields

∂x∗ ∂x∗
gx + gx + gα ≡ 0 (1.40)
∂α ∂α
If we multiply equation 1.40 by λ and combine the result with equa-
tion 1.38 and rearranging terms, we get
∂V ∂x∗ ∂y ∗
= (fx + λgx ) + (fy + λgy ) + fα + λgα = Zα (1.41)
∂α ∂α ∂α
Where Zα is the partial deviative of the Lagrangian function with
respect to α, holding all other variable constant. In this case, the
Langrangian functions serves as the objective function in deriving the
indirect objective function.
While the results in equation 1.41 nicely parallel the unconstrained
case, it is important to note that some of the comparative static results
depend critically on whether the parameters enter only the objective
function or whether they enter only the constraints, or enter both. If a
parameter enters only in the objective function then the comparative
static results are the same as for unconstrained case. However, if the
parameter enters the constraint, the relation

Vαα ≥ fαα

will no longer hold.

8 Further topics in optimization

Interpretation of the Lagrange Multiplier

In the consumer choice problem in chapter 12 we derived the result
that the Lagrange multiplier, λ, represented the change in the value
of the Lagrange function when the consumer’s budget changed. We
loosely interpreted λ as the marginal utility of income. Now let us
derive a more general interpretation of the Lagrange multiplier with
the assistance of the envelope [Link] the problem
Maximize
U = f (x, y) (1.42)
Subject to
c − g(x, y) = 0 (1.43)
where c is a constant. The Lagrangian for this problem is

Z = f (x, y) + λ(c − g(x, y)) (1.44)

The first order equations are

Zx = fx (x, y) − λgx (x, y) = 0

Zy = fy (x, y) − λgy (x, y) = 0 (1.45)
Zλ = c − g(x, y) = 0

From the firs twot equations in (1.45), we get

fx fy
λ= = (1.46)
gx gy
which gives us the condition that the slope of the level curve of the
objective function must equal the slope of the constraint at the opti-
mum.
Equations (1.45) implicitly define the solutions

x = x∗ (c) y = y(c) λ = λ∗ (c) (1.47)

substituting (1.47) back into the Lagrangian yields the mamximum

value function

V (c) = Z ∗ (c) = f (x∗ (c), y ∗ (c)) + λ∗ (c) (c − g(x∗1 (c), y ∗ (c))) (1.48)

differentiating with respect to c yields

∂Z ∗ ∂x∗ ∂y ∗ ∂λ∗ ∗ ∂x∗ ∗ ∂y ∗ ∗ ∂c
= fx +fy +(c − g(x∗ (c), y ∗ (c))) −λ (c)gx −λ (c)gy +λ (c)
∂c ∂c ∂c ∂c ∂c ∂c ∂c
(1.49)
by rearranging we get
∂Z ∗ ∂x∗ ∂y ∗ ∂λ∗
= (fx −λ∗ gx ) +(fy −λ∗ gy ) +(c−g(x∗ , y ∗ )) +λ∗ (1.50)
∂c ∂c ∂c ∂c
Further topics in optimization 9

Note that the three terms in brackets are nothing more than the
first order equations and, at the optimal values of x, y and λ, these
terms are all equal to zero. Therefore this expression simplifies to
∂V (c) ∂Z ∗
= = λ∗ (1.51)
∂c ∂c
Therefore equals the rate of change of the maximum value of the
objective function when c changes (λ is sometimes referred to as the
”shadow price” of c).Note that, in this case, c enters the problem only
through the constraint; it is not an argument of the original objective
function.

Duality and the Envelope Theorem

A consumer’s expenditure function and his indirect utility function are
the minimum and maximum value functions for dual problems. An
expenditure function specifies the minimum expenditure required to
obtain a fixed level of utility given the utility function and the prices
of consumption goods. An indirect utility function specifies the maxi-
mum utility that can be obtained given prices, income and the utility
function.
Let U(x, y) be a utility function in x and y are consumption goods.
The consumer has a budget, B, and faces market prices Px and Py for
goods x and y respectively.
Setting up the Lagrangian:
Z = U(x, y) + λ(B − Px x − Py y) (1.52)
The first order conditions are
Zx = Ux − λPx = 0
Zy = Uy − λPy = 0 (1.53)
Zλ = B − Px X − Py Y = 0
This system of equations implicity defines a solution for xM , y M and
M
λ as a function of the exogenous variables B, Px , Py .
xM = xM (Px , Py , B)
y M = y M (Px , Py , B) (1.54)
λM = λM (Px , Py , B, α)
The solutions to xM and y M are the consumer’s ordinary demand
functions, sometimes called the ”Marshallian” demand functions.3
3
Named after the famous economist Alfred Marshall, known to most economic
students as ”another dead guy.”
10 Further topics in optimization

Substituting the solutions to x∗ and y ∗ into the utility function

yields

U ∗ = U ∗ (xM (B, Px , Py ), y M (B, Px , Py )) = V (B, Px , Py ) (1.55)

Where V is the maximum value function, or indirect utility func-

tion.
Now consider the alternative, or dual, problem for the consumer;
minimize total expenditure on x and y while maintaining a given level
of utility, U ∗ . The Langranian for this problem is

Z = Px x + Py y + λ(U ∗ − U(x, y)) (1.56)

The first order conditions are

Zx = Px − λUx = 0
Zy = Py − λUy = 0 (1.57)
Zλ = U ∗ − U(x, y; α) = 0

This system of equations implicitly define the solutions to xh , y h

and λh
xh = xh (U ∗ , Px , Py )
y h = y h (U ∗ , Px , Py ) (1.58)
h h ∗
λ = λ (U , Px , Py )
xh and y h are the compensated, or ”real income” held constant de-
mand functions. They are commonly referred to as ”Hicksion” demand
functions, hence the h superscript.4
If we compare the first two equations from the first order conditions
in both utility maximization problem and expenditure minimization
problem (Zx , Zy ), we see that both sets can be combined (eliminating
λ) to give us
Px Ux
= (= MRS) (1.59)
Py Uy
This is the tangency condition in which the consumer chooses the
optimal bundle where the slope of the indifference curve equals the
slope of the budget constraint. The tangency condition is identical
for both problems. If the target level of utility in the minimization
problem is set equal to the value of the utility obtained in the solution
to the maximization problem, namely U ∗ , we obtain the following

xM (B, Px , Py ) = xh (U ∗ , Px , Py )
(1.60)
y M (B, Px , Py ) = y h (U ∗ , Px , Py )

or the solution to both the maximization problelm and the mini-

mization problem produce identical values for x and y. However, the
4
Yet another famous, but dead economist, Sir John Hicks.
Further topics in optimization 11

solutions are functions of different exogenous variables so any compar-

ative statics exercises will produce different results.
Substituting xh and y h into the objective function of the minimiza-
tion problem yields

Px xh (Px , Py , U ∗ ) + Py y h (Px , Py , U ∗ ) = E(Px , Py , U ∗ ) (1.61)

where E is the minimum value function or expenditure function.

The duality relationship in this case is

E(Px , Py , U ∗ , α) = B (1.62)

where B is the exogenous budget from the maximization problem.

Finally, it can be shown from the first order conditions of the two
problems that
1
λM = h (1.63)
λ

Roy’s Identity
One application of the envelope theorem is the derivation of Roy’s iden-
tity. Roy’s identity states that the individual consumer’s marshallian
demand function is equal to the ratio of partial derviatives of the max-
imum value function. Substituting the optimal values of xM , y M and
λM into the Lagrangian gives us

V (B, Px , Py ) = U(xM , y M ) + λM (B − Px xM − Py y M ) (1.64)

First differentiate with respect to Px

M
∂V M ∂xM M ∂y M M ∂λ
= (Ux −λ Px ) +(Uy −λ Py ) M
+(B−Px x −Py y ) −λM xM
∂Px ∂Px ∂Px ∂Px
(1.65)

∂V ∂xM ∂y M ∂λM
= (0) + (0) + (0) − λM xM = −λM xM (1.66)
∂Px ∂Px ∂Px ∂Px
Next, differentiate the value function with respect to B
M
∂V M ∂xM M ∂y M M ∂λ
= (Ux −λ Px ) +(Uy −λ Py ) M
+B−Px x −Py y ) +λM
∂B ∂B ∂B ∂B
(1.67)

∂V ∂xM ∂y M ∂λM
= (0) + (0) + (0) + λM = λM (1.68)
∂B ∂B ∂B ∂B
12 Further topics in optimization

Finally, taking the ratio of the two partial derivatives

∂V
∂Px −λM xM
∂V
= M
= xM (1.69)
∂B λ
which is Roy’s identity.

Shephard’s Lemma
Earlier in the chapter an application of the envelope theorem was the
derivation of Hotelling’s Lemma, which states that the partial deriva-
tives of the maximum value of the profit function yields the firm’s fac-
tory demand functions and the supply functions. A similar approach
applied to the expenditure function yields Shepard’s Lemma.
Consider the consumer’s minimization problem. The Lagrangian is
Z = Px x + Py y + λ(U ∗ − U(x, y)) (1.70)
From the first order conditions, the solutions are implicitly defined
xh = xh (Px , Py , U ∗ )
y h = y h (Px , Py , U ∗ ) (1.71)
λh = λh (Px , Py , U ∗ )
Substituting these solutions into the Lagrangian yields the mini-
mum value function
V (Px , Py , U ∗ ) = Px xh + Py y h + λh (U ∗ − U (xh , y h )) (1.72)
The partial derivatives of the value function with respect to Px and
Py are the consumer’s conditional, or Hicksian, demands:
h ∂y h h
∂V
∂Px
= (Px − λh Ux ) ∂P
∂x
x
+ (Py − λh Uy ) ∂P x
∂λ
+ (U ∗ − U (xh , y h )) ∂P x
+ xh
h ∂y h h
∂V ∂x ∂λ
∂Px
= (0) ∂P x
+ (0) ∂P x
+ (0) ∂P x
+ xh = xh
(1.73)
and
h ∂y h h
∂V
∂Py
= (Px − λh Ux ) ∂P
∂x
y
+ (Py − λh Uy ) ∂P y
∂λ
+ (U ∗ − U(xh , y h )) ∂P y
+ yh
h ∂y h h
∂V ∂x
∂Py
= (0) ∂P y
+ (0) ∂P y
+ (0) ∂λ
∂Py
+ yh = yh
(1.74)
Differentiating V with respect to the constraint U∗ yields λh , the
marginal cost of the constraint
h
∂V h ∂xh h ∂y h h h ∂λ
= (Px − λ Ux ) + (Py − λ Uy ) + (U ∗
− U(x , y )) + λh
∂U ∗ ∂U ∗ ∂Py ∂U ∗
∂V ∂xh ∂y h ∂λh
∗
= (0) ∗ + (0) ∗ + (0) ∗ + y h = λh
∂U ∂U ∂U ∂U
Together, these three partial derivatives are Shepard’s Lemma.
Further topics in optimization 13

Example of duality for the consumer choice problem

Utility Maximization
Consider a consumer with the utility function U = xy, who faces a
budget constraint of B = Px xPy y, where all variables are defined as
before.
The choice problem is
Maximize
U = xy (1.75)
Subject to
B = Px xPy y (1.76)
The Lagrangian for this problem is

Z = xy + λ(B − Px xPy y) (1.77)

The first order conditions are

Zx = y − λPx = 0
Zy = x − λPy = 0 (1.78)
Zλ = B − Px x − Py y = 0
Solving the first order conditions yield the following solutions
B B B
xM = 2Px
yM = 2Py
λ= 2Px Py (1.79)

where xM and y M are the consumer’s Marshallian demand func-

tions. Checking second order conditions, the bordered Hessian is
¯ ¯
¯ 1 −Px ¯¯
¯ ¯ ¯ 0
¯H ¯ = ¯ 1 0 −Py ¯¯ = 2Px Py > 0 (1.80)
¯
¯ −Px −Py 0 ¯
Therefore the solution does represent a maximum . Substituting
x and y M into the utility function yields the indirect utility function
M

µ ¶µ ¶
B B B2
V (Px , Py , B) = = (1.81)
2Px 2Py 4Px Py
If we denote the maximum utility by U0 and re-arrange the indirect
utility function to isolate B
B2
= U0 (1.82)
4Px Py
1 1 1 1
B = (4Px Py U0 ) 2 = 2Px2 Py2 U02 = E(Px , Py , U0 ) (1.83)
We have the expenditure function
14 Further topics in optimization

Roy’s Identity Let’s verify Roy’s identity which states

∂V
M ∂Px
x = − ∂V (1.84)
∂B

Taking the partial derivative of V

∂V B2
=− 2 (1.85)
∂Px 4Px Py
and
∂V B
=− (1.86)
∂B Px Py
Taking the negative of the ratio of these two partials
³ 2 ´
∂V B
∂Px 4Px2 Py B
− ∂V = − ³ ´ = = xM (1.87)
∂B
B 2Px
Px Py

Thus we find that Roy’s Identity does hold.

The dual and Shepard’s Lemma

Now consider the dual problem of cost minimization given a fixed level
of utility. Letting U0 denote the target level of utility, the problem is
Minimize
Px x + Py y (1.88)
Subject to
U0 = xy (1.89)
The Lagrangian for the problem is

Z = Px x + Py y + λ(U0 − xy) (1.90)

The first order conditions are

Zx = Px − λy = 0
Zy = Py − λx = 0 (1.91)
Zλ = U0 − xy = 0

Solving the system of equations for x, y and λ

³ ´ 12
Py U0
xh = Px
³ ´ 12
h Px U0 (1.92)
y = Py
³ ´ 12
Px Py
λh = U0
Further topics in optimization 15

where xh and y h are the consumer’s compensated (Hicksian) de-

mand functions. Checking the second order conditions for a minimum
¯ ¯
¯ ¯
¯ ¯ ¯ 0 −λ −y ¯
¯H ¯ = ¯ −λ 0 −x ¯ = −2xyλ < 0 (1.93)
¯ ¯
¯ −y −x 0 ¯

Thus the sufficient conditions for a minimum are satisfied.

Substituting xh and y h into the orginal objective function gives us
the minimum value function, or expenditure function

³ ´ 12 ³ ´ 12
h h Py U0 Px U0
Px x + Py y = Px Px
+ Py Py
1
= (Px Py U0 ) + (Px Py U0 )
2
1
2 (1.94)
1 1 1
2
= 2Px Py U0 2 2

Note that the expenditure function derived here is identical to

the expenditure function obtained by re-arranging the indirect utility
function from the maximization problem.

Shepard’s Lemma We can now test Shepard’s Lemma by differen-

tiating the expenditure function directly.
First, we derive the conditional demand functions
1 1
∂E(Px , Py , U0 ) ∂ ³ 12 12 12 ´ Py2 U02
= 2Px Py U0 = 1 = xh (1.95)
∂Px ∂Px Px 2

and
1 1
∂E(Px , Py , U0 ) ∂ ³ 12 12 12 ´ Py2 U02
= 2Px Py U0 = 1 = yh (1.96)
∂Py ∂Py Py 2

Next, we can find the marginal cost of utility (the Lagrange multi-
plier)

1 1
∂E(Px , Py , U0 ) ∂ ³ 12 12 12 ´ Px2 Py2
0
= 0
2Px Py U0 = 1 = λh (1.97)
∂U ∂U U02

Thus, Shepard’s Lemma holds in this example.

16 Further topics in optimization

Income and Substitution Effects:

The Slutsky Equation
The Traditional Approach
Consider a representative consumer who chooses only two goods: x
and y. The price of both goods are determined in the market and are
therefore exogenous. As well, the consumer’s budget is also exogenously
determined. The consumer choice problem then is
Maximize
U(x, y) (1.98)
Subject to
B = Px X + Py Y (1.99)
The Langrangian function for this optimization problem is

Z = U(x, y) + λ(B − Px x + Py y) (1.100)

The first order conditions yield the following set of simultaneous

equations:
Zλ = B − Px x − Py y = 0
Zx = Ux − λPx = 0 (1.101)
Zy = Uy − λPy = 0
Solving this system will allow us to express the optimal values of the
endogenous variables as implicit functions of the exogenous variables:

λ∗ = λ∗ (Px , Py , B)
x∗ = x∗ (Px , Py , B)
y ∗ = y ∗ (Px , Py , B)

If the bordered Hessian in the present problem is positive

¯ ¯
¯ ¯
¯ ¯ ¯ 0 −Px −Py ¯
¯H ¯ = ¯ −Px Uxx Uxy ¯ = 2Px Py Uxy −Py2 Uxx −Px2 Uyy > 0 (1.102)
¯ ¯
¯ −Py Uyx Uyy ¯

then the value of U will be a maximum.

.
By substituting the optimal values x∗ , y ∗ and λ∗ into the first order
equations, we convert these equations into equilibrium identities:

B − Py x∗ − Py y ∗ ≡ 0
Ux (x∗ , y ∗ ) − λ∗ Px ≡ 0
Uy (x∗ , y ∗ ) − λ∗ Py ≡ 0
Further topics in optimization 17

By taking the total differential of each identity in turn, and noting

that Uxy = Uyx (Young’s Theorem), we then arrive at the linear system
−P xdx∗ − P ydy = x∗ dP x + y ∗ dP y − dB
−P xdλ∗ + Uxx dx∗ + Uxy dy ∗ = λ∗ dP x (1.103)
−P ydλ∗ + Uyx dx∗ + Uyy dy ∗ = λ∗ dP y
Writing these equations in matrix form
⎡ ⎤⎛ ∗ ⎞ ⎡ ∗ ⎤
0 −Px −Py dλ x dPx + y ∗ dPy − dB
⎣ −Px Uxx Uxy ⎦ ⎝ dx∗ ⎠ = ⎣ λ∗ dPx ⎦
∗ ∗
−Py Uyx Uyy dy λ dPy
To study the effect of a change in the budget, let the other exogenous
differentials equal zero (dPx = dPy = 0, dB 6= 0). Then dividing
through by dB, and applying the implicit function theorem, we have
⎡ ⎤⎛ ∗ ⎞ ⎛ ⎞
0 −Px −Py dλ /∂B −1
⎣ −Px Uxx Uxy ⎦ ⎝ dx∗ /∂B ⎠ = ⎝ 0 ⎠ (1.104)
∗
−Py Uyx Uyy dy /∂B 0
The coefficient matrix of this system is the
¯ ¯Jacobian matrix, which
has the same value as the bordered Hessian ¯H ¯ which is positive if the
second order conditions are met. By using Cramer’s rule we can solve
for the following comparative static
¯ ¯
¯ 0 −1 −Py ¯ ¯ ¯ ¯ ¯
∂x ∗
1 ¯¯ ¯ 1 ¯ −Px Uxy ¯ ¯H̄12 ¯
= ¯ ¯ ¯ −Px 0 Uxy ¯¯ = ¯ ¯ ¯¯ ¯= ¯ ¯ ≶0
∂B ¯H ¯ ¯ ¯ ¯H ¯ −Py Uyy ¯ ¯H ¯
−Py 0 Uyy
(1.105)
As before, in the absence of additional information about the rel-
ative magnitudes of Px , Py and the cross partials, Uij , we are un-
able to ascertain the sign of this comparative-static derivative. This
means that the optimal x∗ may increase in the budget, B, depending
on whether it is a normal or inferior good (ambiguous income effect)
Next, we may analyze the effect of a change in Px . Letting dPy =
dB = 0 but keeping dPx 6= 0 and dividing Equation 1.103 by dPx we
obtain
⎡ ⎤⎛ ⎞ ⎡ ∗ ⎤
0 −Px −Py ∂λ∗ /∂Px x
⎣ −Px Uxx Uxy ⎦ ⎝ ∂x∗ /∂Px ⎠ = ⎣ λ∗ ⎦ (1.106)
∗
−Py Uyx Uyy ∂y /∂Px 0
From this, the following comparative static emerges:
¯ ¯
¯ 0 x∗ −Py ¯¯
¯
∂x∗
= H1 ¯¯ −Px λ∗ Uxy ¯¯
∂Px | |¯ ¯
¯ −Py 0 U ¯ yy ¯ ¯
¯ −P U ¯ ∗ ¯ 0 −P ¯ (1.107)
= −x
∗
¯ x xy ¯+ λ ¯ y ¯
|H | ¯ −Py Uyy ¯ |H | ¯ −Py Uyy ¯
|H̄12 | |H̄22 |
= (−x∗ ) H + λ∗ H
| | | |
18 Further topics in optimization

∂x ∗
Note that there are two componants in ( ∂P x
). By comparing the
∂x∗
first term to our previous comparative static ( ∂B ), we see that
¯ ¯ µ ∗¶
¯ 12 ¯
∗ H̄ ∂x
(−x ) ¯ ¯ = (−x )∗
≶0 (1.108)
¯H ¯ ∂B

which can be interpreted as the income effect of a price change.

The second term is the income compensated version of ∂x∗ /∂Px , or the
substitution effect of a price change, which is unambiguously negative:
µ ∗¶ ¯ ¯ ¯ ¯
∂x λ∗ ¯¯ 0 −Py ¯¯ ¯H̄22 ¯ λ∗
∗
= ¯ ¯¯ = λ ¯ ¯ = ¯ ¯ (P 2 ) < 0
∂Px compensated ¯H ¯ −Py Uyy ¯ ¯H ¯ ¯H ¯ y
(1.109)
Hence, we can express Equation 1.107 in the form
µ ∗¶ µ ∗¶
∂x∗ ∂x ∗ ∂x
∗
=− x + (1.110)
∂P ∂B ∂Px compensated
| {z } | {z }
Income Effect Substitution Effect

This result, which decomposes the comparative static derivative

(∂x∗ /∂Px ) into two componants, an income effect and a substitution
effect, is the two-good version of the ”Slutsky Equation.”

Duality and the Alternative Slutsky

From the envelope theorem, we can derive the Slutsky decomposition in
a more succinct manner. Consider first that from the utility maximum
problem we derived solutions for x and y
xM = xM (Px , Py , B)
(1.111)
y M = y M (Px , Py , B)
which were the marshallian demand functions. Substituting these so-
lutions into the utility function yielded the indirect utility function (or
maximum value function)

U ∗ = U(xM (Px , Py , B), y M (Px , Py , B)) = U ∗ (Px , Py , B) (1.112)

which could be rewritten to isolate B and giving us the expenditure

function

B ∗ = B(Px , Py, U ∗ ) (1.113)

Second, from the budget minimization problem we derived the Hick-
sian, or compensated, demand function

x∗ = xh (Px , Py, U ∗ ) (1.114)

Further topics in optimization 19

which, by Shephards lemma, is equivalent to the partial derivative of

the expenditure function with respect to Px :
∂B(Px , Py, U ∗ )
= xc (Px , Py, U ∗ ) (1.115)
∂Px
Thus we know that if the maximum value of utility obtained from

Max U(x, y) + λ(B − Px x − Py y)

is the same value as the exogenous level of utility found in the con-
strained minimization problem

Min Px x + Py y + λ(U0 − U(x, y)) (1.116)

the values of x and y that satisfy the first order conditions of both
problems will be identical, or

xc (Px , Py, U0 ) = xm (Px , Py, B) (1.117)

at the [Link] we subsitiute the expenditure function into xM

in place of the budget, B, we get

xc (Px , Py, U0 ) = xM (Px , Py, B ∗ (Px , Py, U0 )) (1.118)

Differentiate both sides of equation 1.118 with respect to Px

∂xc (Px ,Py, U0 ) ∂xM (Px ,Py, B ∗ (Px ,Py, U0 ))
∂Px
= ∂Px
M B ∗ (Px ,Py, U0 )) ∂B(Px ,Py, U0 ) (1.119)
+ ∂x (Px ,Py,∂B ∂Px

But we know from Shephard’s lemma that

∂B(Px , Py, U0 )
= xc (1.120)
∂Px
substituting equation 1.120 in to equation 1.119 we get
∂xc ∂xM ∂xM
= + xc (1.121)
∂Px ∂Px ∂B
M
Subtract (xc ∂x
∂B
) from both sides gives us

∂xM ∂xM ∂xc

= −xc + (1.122)
∂Px | {z∂B} ∂Px
|{z}
Income effect Substitution effect

If we compare equation (1.122) to equation (1.110) we see that we

have arrived at the identical result. The method of deriving the slut-
sky decomposition through the application of duality and the envelope
theorem is sometimes referred to as the ”instant slutsky”.
20 Further topics in optimization

Problems:
1. A consumer has the following utility function: U (x, y) = x(y +1),
where x and y are quantities of two consumption goods whose
prices are px and py respectively. The consumer also has a budget
of B. Therefore the consumer’s maximization problem is

x(y + 1) + λ(B − px x − py y)

(a) From the first order conditions find expressions for the de-
mand functions. What kind of good is y? In particular what
happens when py > B/2?
(b) Verify that this is a maximum by checking the second or-
der conditions. By substituting x∗ and y ∗ into the utility
function find an expressions for the indirect utility function

U ∗ = U (px , py , B)

and derive an expression for the expenditure function

B ∗ = B(px , py , U ∗ )

(c) This problem could be recast as the following dual problem

Minimize px x + py y subject to U ∗ = x(y + 1)

Find the values of x and y that solve this minimization prob-

lem and show that the values of x and y are equal to the
partial derivatives of the expenditure function, ∂B/∂px and
∂B/∂py respectively.