The Kuhn-Tucker and Envelope Theorems

Peter Ireland∗

ECON 772001 - Math for EconomistsBoston College, Department of Economics

Fall 2020

The Kuhn-Tucker and envelope theorems can be used to characterize the solution toa wide range of constrained optimization problems: static or dynamic, and under perfectforesight or featuring randomness and uncertainty. In addition, these same two resultsprovide foundations for the work on the maximum principle and dynamic programming thatwe will do later on. For both of these reasons, the Kuhn-Tucker and envelope theoremsprovide the starting point for our analysis. Let’s consider each in turn, first in abstractbut somewhat special settings, then applied to some economic examples, and finally in fullgenerality.

1 The Kuhn-Tucker Theorem

References:

Dixit, Chapters 2 and 3.

Simon-Blume, Chapters 18 and 19.

Acemoglu, Appendix A.

Consider a simple constrained optimization problem:

x ∈ R choice variable

F : R→ R objective function, continuously differentiable

c ≥ G(x) constraint, with c ∈ R and G : R→ R, also continuously differentiable.

The problem can be stated as:

maxx

F (x) subject to c ≥ G(x)

∗These lecture notes are licensed under a Creative Commons Attribution-NonCommerical-ShareAlike 4.0International (CC BY-NC-SA 4.0) License. http://creativecommons.org/licenses/by-nc-sa/4.0/

1

This problem is “simple” because it is static and contains no random or stochastic elementsthat would force decisions to be made under uncertainty. This problem is also “simple”because it has a single choice variable and a single constraint. All these simplificationswill make our statement and proof of the Kuhn-Tucker theorem as clean and intuitiveas possible. But the results can be generalized along all of these dimensions and,throughout the semester, we will work through examples that do so.

Probably the easiest way to solve this problem is via the method of Lagrange multipliers.The mathematical foundations that allow for the application of this method are givento us by Lagrange’s Theorem or, in its most general form, the Kuhn-Tucker Theorem.

To prove this theorem, begin by defining the Lagrangian:

L(x, λ) = F (x) + λ[c−G(x)]

for any x ∈ R and λ ∈ R.

Theorem (Kuhn-Tucker) Suppose that x∗ maximizes F (x) subject to c ≥ G(x), whereF and G are both continuously differentiable, and suppose that G′(x∗) 6= 0. Thenthere exists a value λ∗ of λ such that x∗ and λ∗ satisfy the following four conditions:

L1(x∗, λ∗) = F ′(x∗)− λ∗G′(x∗) = 0, (1)

L2(x∗, λ∗) = c−G(x∗) ≥ 0, (2)

λ∗ ≥ 0, (3)

andλ∗[c−G(x∗)] = 0. (4)

Proof Consider two possible cases, depending on whether or not the constraint is bindingat x∗.

Case 1: Nonbinding Constraint.

If c > G(x∗), then let λ∗ = 0. Clearly, (2)-(4) are satisfied, so it only remains to showthat (1) must hold. With λ∗ = 0, (1) holds if and only if

F ′(x∗) = 0. (5)

We can show that (5) must hold using a proof by contradiction. Suppose thatinstead of (5), it turns out that

F ′(x∗) < 0.

Then, by the continuity of F and G, there must exist an ε > 0 such that

F (x∗ − ε) > F (x∗) and c > G(x∗ − ε).

2

But this result contradicts the assumption that x∗ maximizes F (x) subject toc ≥ G(x). Similarly, if it turns out that

F ′(x∗) > 0,

then by the continuity of F and G there must exist an ε > 0 such that

F (x∗ + ε) > F (x∗) and c > G(x∗ + ε),

But, again, this result contradicts the assumption that x∗ maximizes F (x) subjectto c ≥ G(x). This establishes that (5) must hold, completing the proof for case 1.

Case 2: Binding Constraint.

If c = G(x∗), then let λ∗ = F ′(x∗)/G′(x∗). This is possible, given the assumptionthat G′(x∗) 6= 0. Clearly, (1), (2), and (4) are satisfied, so it only remains to showthat (3) must hold. With λ∗ = F ′(x∗)/G′(x∗), (3) holds if and only if

F ′(x∗)/G′(x∗) ≥ 0. (6)

We can show that (6) must hold using a proof by contradiction. Suppose thatinstead of (6), it turns out that

F ′(x∗)/G′(x∗) < 0.

One way that this can happen is if F ′(x∗) > 0 and G′(x∗) < 0. But if theseconditions hold, then the continuity of F and G implies the existence of an ε > 0such that

F (x∗ + ε) > F (x∗) and c = G(x∗) > G(x∗ + ε),

which contradicts the assumption that x∗ maximizes F (x) subject to c ≥ G(x).And if, instead, F ′(x∗)/G′(x∗) < 0 because F ′(x∗) < 0 and G′(x∗) > 0, then thecontinuity of F and G implies the existence of an ε > 0 such that

F (x∗ − ε) > F (x∗) and c = G(x∗) > G(x∗ − ε),

which again contradicts the assumption that x∗ maximizes F (x) subject to c ≥G(x). This establishes that (6) must hold, completing the proof for case 2.

Notes:

a) The theorem can be extended to handle cases with more than one choice variableand more than one constraint: see Dixit, Simon-Blume, Acemoglu, or section 4.1of the notes below.

b) The extra assumption that G′(x∗) 6= 0 is needed to guarantee the existence ofa multiplier λ∗ satisfying (1)-(4) in the case where the constraint binds at theoptimum, so that c = G(x∗). This extra assumption is called the “constraintqualification.” In economic applications, this constraint qualification is almostalways satisfied. But it is important to remember that, in cases where the con-straint qualification fails to hold, it may be impossible to find a value of λ∗ that,together with the the value x∗ that solves the constrained optimization problem,satisfies (1)-(4).

3

c) Equations (1)-(4) are necessary conditions: If x∗ is a solution to the optimizationproblem, then there exists a λ∗ such that (1)-(4) must hold. But (1)-(4) are notsufficient conditions: if x∗ and λ∗ satisfy (1)-(4), it does not follow automaticallythat x∗ is a solution to the optimization problem.

As an example that illustrates point (b), consider the problem:

maxx

ex − 2

1 + x2subject to 1 ≥ x.

With the Lagrangian defined as

L(x, λ) = ex − 2

1 + x2+ λ(1− x),

the Kuhn-Tucker conditions are

L1(x∗, λ∗) = ex∗

+4x∗

[1 + (x∗)2]2− λ∗ = 0,

L2(x∗, λ∗) = 1− x∗ ≥ 0,

λ∗ ≥ 0,

andλ∗(1− x∗) = 0.

These conditions are satisfied when x∗ = 1 and λ∗ = e + 1, which corresponds to thesolution to the problem, but they are also satisfied when x∗ = −0.2205 and λ∗ = 0,which corresponds instead to the solution to the minimization problem

minxex − 2

1 + x2subject to 1 ≥ x.

But despite point (b) listed above, the Kuhn-Tucker theorem is extremely useful in practice.Suppose that we are looking for the solution x∗ to the constrained optimization problem

maxx

F (x) subject to c ≥ G(x).

The theorem tells us that if we form the Lagrangian

L(x, λ) = F (x) + λ[c−G(x)],

then x∗ and the associated λ∗ must satisfy the first-order condition (FOC) obtainedby differentiating L by x and setting the result equal to zero:

L1(x∗, λ∗) = F ′(x∗)− λ∗G′(x∗) = 0, (1)

In addition, we know that x∗ must satisfy the constraint:

c ≥ G(x∗). (2)

4

We know that the Lagrange multiplier λ∗ must be nonnegative:

λ∗ ≥ 0. (3)

And finally, we know that the complementary slackness condition

λ∗[c−G(x∗)] = 0, (4)

must hold: If λ∗ > 0, then the constraint must bind; if the constraint does not bind,then λ∗ = 0.

In searching for the value of x that solves the constrained optimization problem, we onlyneed to consider values of x∗ that satisfy (1)-(4). In the example from above, forinstance, it is straightforward to compare the values of the objective function whenx = 1 and x = −0.2205 to conclude that the solution to the optimization problem isx∗ = 1.

Furthermore, in many economic applications, the objective function F (x) will be con-cave and the function G(x) entering into the constraint will be convex. Under theseadditional assumptions, the Kuhn-Tucker conditions (1)-(4) are sufficient as well asnecessary.

To see this, let x∗ and λ∗ satisfy (1)-(4), and let x be any other value of the choice variablethat satisfies the constraint c ≥ G(x). Since F is concave, it satisfies by definition

ωF (x) + (1− ω)F (x∗) ≤ F [ωx+ (1− ω)x∗]

for all ω ∈ [0, 1]. Let ω > 0, and rewrite this inequality as

F (x)− F (x∗) ≤ F [x∗ + ω(x− x∗)]− F (x∗)

ωor

F (x)− F (x∗) ≤{F [x∗ + ω(x− x∗)]− F (x∗)

ω(x− x∗)

}(x− x∗).

Taking the limit as ω → 0 of the term inside brackets on the right-hand side of thislast inequality yields

F (x)− F (x∗) ≤ F ′(x∗)(x− x∗)or, since F ′(x∗) = λ∗G′(x∗) by (1),

F (x)− F (x∗) ≤ λ∗G′(x∗)(x− x∗).

Observe next that since G is convex, it satisfies by definition

ωG(x) + (1− ω)G(x∗) ≥ G[ωx+ (1− ω)x∗]

for all ω ∈ [0, 1]. Let ω > 0, rewrite this inequality as

G(x)−G(x∗) ≥{G[x∗ + ω(x− x∗)]−G(x∗)

ω(x− x∗)

}(x− x∗),

and take the limit as ω → 0 of the term inside brackets on the right-hand side to obtain

G(x)−G(x∗) ≥ G′(x∗)(x− x∗).

5

Since λ∗ ≥ 0 by (3),F (x)− F (x∗) ≤ λ∗G′(x∗)(x− x∗).

andG(x)−G(x∗) ≥ G′(x∗)(x− x∗).

can be combined to yield

F (x)− F (x∗) ≤ λ∗[G(x)−G(x∗)]

or, since (4) implies λ∗G(x∗) = λ∗c,

F (x)− F (x∗) ≤ λ∗[G(x)− c].

Since the constraint on x requires c ≥ G(x), this last inequality confirms that

F (x)− F (x∗) ≤ 0

or F (x∗) ≥ F (x) for all x such that c ≥ G(x). That is, x∗ must solve the constrainedoptimization problem.

For extensions of this proof of sufficiency to the more general case where there is more thanone choice variable and more than one constraint, see Simon and Blume (Theorem21.22, pp.532-533) or Acemoglu (Theorem A.29, pp.911-913).

Two related pieces of terminology:

a) When F is concave and G is convex, the problem

maxx

F (x) subject to c ≥ G(x),

is called a “concave program.”

b) Simon and Blume’s Theorem 21.22 and Acemoglu’s Theorem A.29 also show thatunder these extra assumptions, (x∗, λ∗) is a “saddle point” of L(x, λ): x∗ maxi-mizes L(x, λ∗), while λ∗ minimizes L(x∗, λ) subject to the constraint that λ ≥ 0.Intuitively, (1) then appears as a first-order condition for the problem of maxi-mizing L with respect to x, while (2) is like a first-order condition for minimizingL with respect to λ.

Thus, point (b) gives us some intuition about how the Kuhn-Tucker theorem works. Ituses the Lagrangian to turn a constrained optimization problem into an unconstrainedproblem, where the solution for x∗ is a critical point of L(x, λ∗) rather than a criticalpoint of F (x). And when the extra assumptions made in concave programming areadopted, x∗ is not only a critical point of L(x, λ∗), it also maximizes L(x, λ∗).

As an example illustrating the saddle-point property, consider the problem

maxx

ln(x) subject to 1 ≥ x.

6

Since the logarithmic objective function is strictly increasing, we know that x∗ = 1 isthe solution. Forming the Lagrangian

L(x, λ) = ln(x) + λ(1− x)

and taking the first-order condition

L1(x∗, λ∗) =1

x∗− λ∗ = 0

reveals that the associated value of λ is λ∗ = 1. Substituting this value for λ back intothe Lagrangian

L(x, λ∗) = ln(x) + 1− x,we can see that L(x, λ∗) is maximized with x∗ = 1.

Note, however, that in the general case where F (x) need not be concave, x∗ will always bea critical point of the Lagrangian – that is, it will satisfy the first-order condition (1)– but x∗ need not maximize L(x, λ). An example of how this can happen is given byDixit’s example 7.2 (p.103). Consider the problem

maxx

ex subject to 1 ≥ x.

Since the exponential objective function is strictly increasing, we know again that thisproblem is solved with x∗ = 1. Forming the Lagrangian and taking the first-ordercondition

L1(x∗, λ∗) = ex∗ − λ∗ = 0,

shows that the associated value of λ∗ is e. But λ∗ = e implies that

L(x, λ∗) = ex + λ∗(1− x) = ex + e(1− x) = ex − ex+ e,

and since ex− ex grows without bound as either x→∞ or x→ −∞, x∗ = 1 does notmaximize the Lagrangian given λ∗. Hence, (x∗, λ∗) is not a saddle-point of L, since theobjective function from the original problem is convex, not concave.

One final note:

Our general constraint, c ≥ G(x), nests as a special case the nonnegativity constraintx ≥ 0, obtained by setting c = 0 and G(x) = −x.

So nonnegativity constraints can be introduced into the Lagrangian in the same wayas all other constraints. If we consider, for example, the extended problem

maxx

F (x) subject to c ≥ G(x) and x ≥ 0,

then we can introduce a second multiplier µ, form the Lagrangian as

L(x, λ, µ) = F (x) + λ[c−G(x)] + µx,

and write the first order condition for the optimal x∗ as

L1(x∗, λ∗, µ∗) = F ′(x∗)− λ∗G′(x∗) + µ∗ = 0. (1′)

In addition, analogs to our earlier conditions (2)-(4) must also hold for the secondconstraint: x∗ ≥ 0, µ∗ ≥ 0, and µ∗x∗ = 0.

7

Kuhn and Tucker’s original statement of the theorem, however, does not incorporatenonnegativity constraints into the Lagrangian. Instead, even with the additionalnonnegativity constraint x ≥ 0, they continue to define the Lagrangian as

L(x, λ) = F (x) + λ[c−G(x)].

If this case, the first order condition for x∗ must be modified to read

L1(x∗, λ∗) = F ′(x∗)− λ∗G′(x∗) ≤ 0,with equality if x∗ > 0. (1′′)

Of course, in (1′), µ∗ ≥ 0 in general and µ∗ = 0 if x∗ > 0. So a close inspection revealsthat these two approaches to handling nonnegativity constraints lead in the endto the same results.

2 The Envelope Theorem

References:

Dixit, Chapter 5.

Simon-Blume, Chapter 19.

Acemoglu, Appendix A.

In our discussion of the Kuhn-Tucker theorem, we considered an optimization problem ofthe form

maxx

F (x) subject to c ≥ G(x)

Now, let’s generalize the problem by allowing the functions F and G to depend on aparameter θ ∈ R. The problem can now be stated as

maxx

F (x, θ) subject to c ≥ G(x, θ)

For this problem, define the maximum value function V : R→ R as

V (θ) = maxx

F (x, θ) subject to c ≥ G(x, θ)

Note that evaluating V requires a two-step procedure:

First, given θ, find the value of x∗ that solves the constrained optimization problem.

Second, substitute this value of x∗, together with the given value of θ, into the objec-tive function to obtain

V (θ) = F (x∗, θ)

Now suppose that we want to investigate the properties of this function V . Suppose, inparticular, that we want to take the derivative of V with respect to its argument θ.

8

As the first step in evaluating V ′(θ), consider solving the constrained optimization problemfor any given value of θ by setting up the Lagrangian

L(x, λ) = F (x, θ) + λ[c−G(x, θ)]

We know from the Kuhn-Tucker theorem that the solution x∗ to the optimization problemand the associated value of the multiplier λ∗ must satisfy the complementary slacknesscondition:

λ∗[c−G(x∗, θ)] = 0

Use this last result to rewrite the expression for V as

V (θ) = F (x∗, θ) = F (x∗, θ) + λ∗[c−G(x∗, θ)]

So suppose that we tried to calculate V ′(θ) simply by differentiating both sides of thisequation with respect to θ:

V ′(θ) = F2(x∗, θ)− λ∗G2(x∗, θ).

But, in principle, this formula may not be correct. The reason is that x∗ and λ∗ willthemselves depend on the parameter θ, and we must take this dependence into accountwhen differentiating V with respect to θ.

However, the envelope theorem tells us that our formula for V ′(θ) is, in fact, correct. Thatis, the envelope theorem tells us that we can ignore the dependence of x∗ and λ∗ on θin calculating V ′(θ).

To see why, for any θ, let x∗(θ) denote the solution to the problem: max F (x, θ) subject toc ≥ G(x, θ), and let λ∗(θ) be the associated Lagrange multiplier.

Theorem (Envelope) Let F and G be continuously differentiable functions of x and θ.For any given θ, let x∗(θ) maximize F (x, θ) subject to c ≥ G(x, θ), and let λ∗(θ) bethe associated value of the Lagrange multiplier. Suppose, further, that x∗(θ) and λ∗(θ)are also continuously differentiable functions, and that the constraint qualificationG1[x∗(θ), θ] 6= 0 holds for all values of θ. Then the maximum value function defined by

V (θ) = maxx

F (x, θ) subject to c ≥ G(x, θ)

satisfiesV ′(θ) = F2[x∗(θ), θ]− λ∗(θ)G2[x∗(θ), θ]. (7)

Proof The Kuhn-Tucker theorem tells us that for any given value of θ, x∗(θ) and λ∗(θ)must satisfy

L1[x∗(θ), λ∗(θ)] = F1[x∗(θ), θ]− λ∗(θ)G1[x∗(θ), θ] = 0, (1)

andλ∗(θ){c−G[x∗(θ), θ]} = 0. (4)

9

In light of (4),V (θ) = F [x∗(θ), θ] = F [x∗(θ), θ] + λ∗(θ){c−G[x∗(θ), θ]}

Differentiating both sides of this expression with respect to θ yields

V ′(θ) = F1[x∗(θ), θ]x∗′(θ) + F2[x∗(θ), θ]

+λ∗′(θ){c−G[x∗(θ), θ]} − λ∗(θ)G1[x∗(θ), θ]x∗′(θ)− λ∗(θ)G2[x∗(θ), θ]

which shows that, in principle, we must take the dependence of x∗ and λ∗ on θ intoaccount when calculating V ′(θ).

Note, however, that

V ′(θ) = {F1[x∗(θ), θ]− λ∗(θ)G1[x∗(θ), θ]}x∗′(θ)+F2[x∗(θ), θ] + λ∗′(θ){c−G[x∗(θ), θ]} − λ∗(θ)G2[x∗(θ), θ],

which by (1) reduces to

V ′(θ) = F2[x∗(θ), θ] + λ∗′(θ){c−G[x∗(θ), θ]} − λ∗(θ)G2[x∗(θ), θ]

Thus, it only remains to show that

λ∗′(θ){c−G[x∗(θ), θ]} = 0 (8)

Clearly, (8) holds for any θ such that the constraint is binding.

For θ such that the constraint is not binding, (4) implies that λ∗(θ) must equal zero.Furthermore, by the continuity of G and x∗, if the constraint does not bind at θ, thereexists an ε∗ > 0 such that the constraint does not bind for all θ + ε with ε∗ > |ε|.Hence, (4) also implies that λ∗(θ + ε) = 0 for all ε∗ > |ε|. Using the definition of thederivative

λ∗′(θ) = limε→0

λ∗(θ + ε)− λ∗(θ)ε

= limε→0

0

ε= 0,

it once again becomes apparent that (8) must hold.

Thus,V ′(θ) = F2[x∗(θ), θ]− λ∗(θ)G2[x∗(θ), θ]

as claimed in the theorem.

Once again, this theorem is useful because it tells us that we can ignore the dependence ofx∗ and λ∗ on θ in calculating V ′(θ).

And once again, the theorem can be extended to apply in more general settings: see Dixit,Simon-Blume, Acemoglu, or section 4.2 of the notes below.

10

It is worth noting that the assumptions required by the envelope theorem are more restric-tive than those required by the Kuhn-Tucker theorem. For example, the statementof the Kuhn-Tucker theorem makes reference to the value x∗ of x that solves theconstrained optimization problem, thereby assuming implicitly that a solution to theproblem exists. But the envelope theorem requires that x∗ be depicted as a function ofθ, implying that the solution must not only exist but be unique as well. Further, thesolution must vary smoothly with θ, so that x∗ and λ∗ can be written as continuouslydifferentiable functions of the parameter.

Although most statements of the envelope theorem assume directly that x∗(θ) and λ∗(θ)are continuously differentiable, it is also possible to impose restrictions on F (x, θ) andG(x, θ) that guarantee this. Assume, for example, that the constraint always binds atthe optimum, so that by the Kuhn-Tucker theorem, x∗(θ) and λ∗(θ) must satisfy

L1[x∗(θ), λ∗(θ)] = F1[x∗(θ), θ]− λ∗(θ)G1[x∗(θ), θ] = 0

andL2[x∗(θ), λ∗(θ)] = λ∗(θ){c−G[x∗(θ), θ]} = 0.

If F and G are two times continuously differentiable in x and θ, and if the matrix ofsecond derivatives of the Lagrangian[

L11[x∗(θ), λ∗(θ)] L12[x∗(θ), λ∗(θ)]L21[x∗(θ), λ∗(θ)] L22[x∗(θ), λ∗(θ)]

]is nonsingular, then the implicit function theorem (Simon and Blume, Theorem 15.7,pp.355-356) will imply that x∗(θ) and λ∗(θ) exist and are continuously differentiable.

But what is the intuition for why the envelope theorem holds? To obtain some intuition,begin by considering the simpler, unconstrained optimization problem:

maxx

F (x, θ),

where x is the choice variable and θ is the parameter.

Associated with this unconstrained problem, define the maximum value function in thesame way as before:

V (θ) = maxx

F (x, θ).

To evaluate V for any given value of θ, use the same two-step procedure as before. First,find the value x∗(θ) that solves the unconstrained maximization problem for that valueof θ. Second, substitute that value of x back into the objective function to obtain

V (θ) = F [x∗(θ), θ].

Now differentiate both sides of this expression through by θ, carefully taking the dependenceof x∗ on θ into account:

V ′(θ) = F1[x∗(θ), θ]x∗′(θ) + F2[x∗(θ), θ].

11

But, if x∗(θ) is the value of x that maximizes F given θ, we know that x∗(θ) must be acritical point of F :

F1[x∗(θ), θ] = 0.

Hence, for the unconstrained problem, the envelope theorem implies that

V ′(θ) = F2[x∗(θ), θ],

so that, again, we can ignore the dependence of x∗ on θ in differentiating the maximumvalue function. And this result holds not because x∗ fails to depend on θ: to thecontrary, in fact, x∗ will typically depend on θ through the function x∗(θ). Instead, theresult holds because since x∗ is chosen optimally, x∗(θ) is a critical point of F given θ.

Now return to the constrained optimization problem

maxx

F (x, θ) subject to c ≥ G(x, θ)

and define the maximum value function as before:

V (θ) = maxx

F (x, θ) subject to c ≥ G(x, θ).

The envelope theorem for this constrained problem tells us that we can also ignore thedependence of x∗ on θ when differentiating V with respect to θ, but only if we start byadding the complementary slackness condition to the maximized objective function tofirst obtain

V (θ) = F [x∗(θ), θ] + λ∗(θ){c−G[x∗(θ), θ]}.

In taking this first step, we are actually evaluating the entire Lagrangian at the optimum,instead of just the objective function. We need to take this first step because for theconstrained problem, the Kuhn-Tucker condition (1) tells us that x∗(θ) is a criticalpoint, not of the objective function by itself, but of the entire Lagrangian formed byadding the product of the multiplier and the constraint to the objective function.

And what gives the envelope theorem its name? The “envelope” theorem refers to a geo-metrical presentation of the same result that we’ve just worked through.

To see where that geometrical interpretation comes from, consider again the simpler, un-constrained optimization problem:

maxx

F (x, θ),

where x is the choice variable and θ is a parameter.

Following along with our previous notation, let x∗(θ) denote the solution to this problemfor any given value of θ, so that the function x∗(θ) tells us how the optimal choice ofx depends on the parameter θ.

12

Also, continue to define the maximum value function V in the same way as before:

V (θ) = maxx

F (x, θ).

Now let θ1 denote a particular value of θ, and let x1 denote the optimal value of x associatedwith this particular value θ1. That is, let

x1 = x∗(θ1).

After substituting this value of x1 into the function F , we can think about how F (x1, θ)varies as θ varies—that is, we can think about F (x1, θ) as a function of θ, holding x1

fixed.

In the same way, let θ2 denote another particular value of θ, with θ2 > θ1 let’s say. Andfollowing the same steps as above, let x2 denote the optimal value of x associated withthis particular value θ2, so that

x2 = x∗(θ2).

Once again, we can hold x2 fixed and consider F (x2, θ) as a function of θ.

The geometrical presentation of the envelope theorem can be derived by thinking about theproperties of these three functions of θ: V (θ), F (x1, θ), and F (x2, θ).

One thing that we know about these three functions is that for θ = θ1:

V (θ1) = F (x1, θ1) > F (x2, θ1),

where the first equality and the second inequality both follow from the fact that, bydefinition, x1 maximizes F (x, θ1) by choice of x.

Another thing that we know about these three functions is that for θ = θ2:

V (θ2) = F (x2, θ2) > F (x1, θ2),

because again, by definition, x2 maximizes F (x, θ2) by choice of x.

On a graph, these relationships imply that:

At θ1, V (θ) coincides with F (x1, θ), which lies above F (x2, θ).

At θ2, V (θ) coincides with F (x2, θ), which lies above F (x1, θ).

And we could find more and more values of V by repeating this procedure for moreand more specific values of θi, i = 1, 2, 3, ....

In other words:

V (θ) traces out the “upper envelope” of the collection of functions F (xi, θ), formedby holding xi = x∗(θi) fixed and varying θ.

13

Moreover, V (θ) is tangent to each individual function F (xi, θ) at the value θi of θ forwhich xi is optimal, or equivalently:

V ′(θ) = F2[x∗(θ), θ],

which is the same analytical result that we derived earlier for the unconstrainedoptimization problem.

If, for example,F (x, θ) = −(x− θ)2 + θ2 = −x2 + 2xθ,

thenV (θ) = max

x−(x− θ)2 + θ2 = θ2,

since, in this case, x∗(θ) = θ for all values of θ.

The figure below sets θ1 = 2 and θ2 = 7; hence x1 = 2 and x2 = 7, then plots

F (x1, θ) = −4 + 4θ,

F (x2, θ) = −49 + 14θ,

andV (θ) = θ2

to show howV (θ1) = F (x1, θ1) > F (x2, θ1) at θ1 = 2,

andV (θ2) = F (x2, θ2) > F (x1, θ2) at θ2 = 7,

and how, more generally, V (θ) traces out the upper envelope of the family of functionsF (xi, θ), where each xi maximizes F (x, θ) for some value θi of θ.

14

To generalize these arguments so that they apply to the constrained optimization problem

maxx

F (x, θ) subject to c ≥ G(x, θ),

simply use the fact that in many cases (as when F is concave and G is convex) thevalue x∗(θ) that solves the constrained optimization problem for any given value of θalso maximizes the Lagrangian function

L(x, λ, θ) = F (x, θ) + λ[c−G(x, θ)],

so that

V (θ) = maxx

F (x, θ) subject to c ≥ G(x, θ)

= maxx

L(x, λ, θ)

Now just replace the function F with the function L in working through the argumentsfrom above to conclude that

V ′(θ) = L3[x∗(θ), λ∗(θ), θ] = F2[x∗(θ), θ]− λ∗(θ)G2[x∗(θ), θ],

which is again the same result that we derived before for the constrained optimizationproblem.

Note that in the figure, the maximum value function V (θ) is convex, with a strictly positivesecond derivative, whereas the functions F (x1, θ) and F (x2, θ) are both linear. Ingeneral, it can be shown that there is a sense in which the maximum value functionV (θ) will always be “more convex” or “less concave” than each member of the familyof functions F (xi, θ).

To see this, consider the two functions V (θ) and F (x1, θ) as they were defined for the generalunconstrained problem

maxx

F (x, θ).

Second order Taylor approximations imply that for values of θ sufficiently close to θ1:

V (θ) ≈ V (θ1) + V ′(θ1)(θ − θ1) + (1/2)V ′′(θ1)(θ − θ1)2

andF (x1, θ) ≈ F (x1, θ1) + F2(x1, θ1)(θ − θ1) + (1/2)F22(x1, θ1)(θ − θ1)2.

However,V (θ1) = F (x1, θ1),

since x1 = x∗(θ1), andV ′(θ1) = F2(x1, θ1),

by the envelope theorem itself. Since the definition of V implies that

V (θ) ≥ F (x1, θ)

15

for all values of θ, the Taylor approximations imply

V ′′(θ1) ≥ F22(x1, θ1),

so that, more specifically, the second derivative of the maximum value function willalways be larger than the second derivatives of the functions F (xi, θ).

Finally, to begin getting a feel for the usefulness of the envelope theorem, consider a pre-liminary economic example. Suppose that a firm hires n workers, paying each thecompetitive real wage w (real wage, so that we don’t have to consider separately theprice of output), in order to produce y units of output according to the productionfunction

nα ≥ y,

where α lies between zero and one: 0 < α < 1.

For simplicity, let’s depict the firm as solving the unconstrained optimization problem

maxn

nα − wn,

with first-order conditionα(n∗)α−1 − w = 0

that leads to the labor demand curve

n∗(w) =

(1

α

) 1α−1

w1

α−1 .

Next, define the maximum value function

V (w) = maxn

nα − wn,

so thatV (w) = [n∗(w)]α − wn∗(w).

The envelope theorem immediately implies that

V ′(w) = −n∗(w) = −(

1

α

) 1α−1

w1

α−1 .

Suppose instead we substitute our earlier solution for n∗(w) into the expression for V (w):

V (w) = [n∗(w)]α − wn∗(w)

=

[(1

α

) 1α−1

w1

α−1

]α− w

(1

α

) 1α−1

w1

α−1

=

(1

α

) αα−1

wαα−1 −

(1

α

) 1α−1

wαα−1

=

[(1

α

) αα−1

−(

1

α

) 1α−1

]w

αα−1 .

16

Now differentiate with respect to w and simplify to get

V ′(w) =

(α

α− 1

)[(1

α

) αα−1

−(

1

α

) 1α−1

]w

1α−1

=

(α

α− 1

)(1

α

) 1α−1(

1

α− 1

)w

1α−1

= −(

1

α

) 1α−1

w1

α−1 ,

exactly as we found, much more quickly, using the envelope theorem.

In terms of its economics, this example shows that when the firm faces a higher wage, therewill be two effects on its profits.

The direct effect: it must pay each of its workers a higher wage, so profits fall by−n∗(w).

The indirect effect: it can try to mitigate the direct effect by hiring fewer workers.

The envelope theorem says that since the firm has already chosen n∗ optimally, thefirst-order effect on profits of adjusting this decision in response to an arbitrarily smallchange in the wage is zero. Only the direct effect remains: V ′(w) = −n∗(w).

3 Three Examples

3.1 Utility Maximization

A consumer has a utility function defined over consumption of two goods: U(c1, c2)

Prices: p1 and p2

Income: I

Budget constraint: I ≥ p1c1 + p2c2 = G(c1, c2)

The consumer’s problem is:

maxc1,c2

U(c1, c2) subject to I ≥ p1c1 + p2c2

The Kuhn-Tucker theorem tells us that if we set up the Lagrangian:

L(c1, c2, λ) = U(c1, c2) + λ(I − p1c1 − p2c2)

Then the optimal consumptions c∗1 and c∗2 and the associated multiplier λ∗ must satisfy theFOC:

L1(c∗1, c∗2, λ∗) = U1(c∗1, c

∗2)− λ∗p1 = 0

andL2(c∗1, c

∗2, λ∗) = U2(c∗1, c

∗2)− λ∗p2 = 0

17

Move the terms with minus signs to the other side, and divide the first of these FOC bythe second to obtain

U1(c∗1, c∗2)

U2(c∗1, c∗2)

=p1

p2

,

which is just the familiar condition that says that the optimizing consumer should setthe slope of his or her indifference curve, the marginal rate of substitution, equal tothe slope of his or her budget constraint, the ratio of prices.

Now consider I as one of the model’s parameters, and let the functions c∗1(I), c∗2(I), andλ∗(I) describe how the optimal choices c∗1 and c∗2 and the associated value λ∗ of themultiplier depend on I.

In addition, define the maximum value function as

V (I) = maxc1,c2

U(c1, c2) subject to I ≥ p1c1 + p2c2

The Kuhn-Tucker theorem tells us that

λ∗(I)[I − p1c∗1(I)− p2c

∗2(I)] = 0

and hence

V (I) = U [c∗1(I), c∗2(I)] = U [c∗1(I), c∗2(I)] + λ∗(I)[I − p1c∗1(I)− p2c

∗2(I)].

The envelope theorem tells us that we can ignore the dependence of c∗1 and c∗2 on I incalculating

V ′(I) = λ∗(I),

which gives us an interpretation of the multiplier λ∗ as the marginal utility of income.

3.2 Cost Minimization

The Kuhn-Tucker and envelope conditions can also be used to study constrained minimiza-tion problems.

Consider a firm that produces output y using capital k and labor l, according to thetechnology described by

f(k, l) ≥ y.

r = rental rate for capital

w = wage rate

Suppose that the firm takes its output y as given, and chooses inputs k and l to minimizecosts. Then the firm solves

mink,l

rk + wl subject to f(k, l) ≥ y

18

If we set up the Lagrangian as

L(k, l, λ) = rk + wl − λ[f(k, l)− y],

where the term involving the multiplier λ is subtracted rather than added in the case ofa minimization problem, the Kuhn-Tucker conditions (1)-(4) continue to apply, exactlyas before.

Thus, according to the Kuhn-Tucker theorem, the optimal choices k∗ and l∗ and the asso-ciated multiplier λ∗ must satisfy the FOC:

L1(k∗, l∗, λ∗) = r − λ∗f1(k∗, l∗) = 0 (9)

andL2(k∗, l∗, λ∗) = w − λ∗f2(k∗, l∗) = 0 (10)

Move the terms with minus signs over to the other side, and divide the first FOC by thesecond to obtain

f1(k∗, l∗)

f2(k∗, l∗)=r

w,

which is another familiar condition that says that the optimizing firm chooses factorinputs so that the marginal rate of substitution between inputs in production equalsthe ratio of factor prices.

Now suppose that the constraint binds, as it usually will:

y = f(k∗, l∗) (11)

Then (9)-(11) represent 3 equations that determine the three unknowns k∗, l∗, and λ∗ asfunctions of the model’s parameters r, w, and y. In particular, we can think of thefunctions

k∗ = k∗(r, w, y)

andl∗ = l∗(r, w, y)

as demand curves for capital and labor: strictly speaking, they are conditional (on y)factor demand functions.

Now define the minimum cost function as

C(r, w, y) = mink,l

rk + wl subject to f(k, l) ≥ y

= rk∗(r, w, y) + wl∗(r, w, y)

= rk∗(r, w, y) + wl∗(r, w, y)− λ∗(r, w, y){f [k∗(r, w, y), l∗(r, w, y)]− y}

The envelope theorem tells us that in calculating the derivatives of the cost function, wecan ignore the dependence of k∗, l∗, and λ∗ on r, w, and y.

19

Hence:C1(r, w, y) = k∗(r, w, y),

C2(r, w, y) = l∗(r, w, y),

andC3(r, w, y) = λ∗(r, w, y).

The first two of these equations are statements of Shephard’s lemma; they tell us thatthe derivatives of the cost function with respect to factor prices coincide with theconditional factor demand curves. The third equation gives us an interpretation of themultiplier λ∗ as a measure of the marginal cost of increasing output.

Thus, our two examples illustrate how we can apply the Kuhn-Tucker and envelope theoremsin specific economic problems.

The two examples also show how, in the context of specific economic problems, it is oftenpossible to attach an economic interpretation to the multiplier λ∗.

3.3 Le Chatelier’s Principle

For the next example, extend the previous cost minimization problem by introducing athird input:

m = materials input

q = price of materials

Define the minimum cost function

C(r, w, q, y) = mink,l,m

rk + wl + qm subject to f(k, l,m) ≥ y.

Then Shephard’s lemma implies

C1(r, w, q, y) = k∗(r, w, q, y),

C2(r, w, q, y) = l∗(r, w, q, y),

andC3(r, w, q, y) = m∗(r, w, q, y),

where k∗(r, w, q, y), l∗(r, w, q, y), and m∗(r, w, q, y) are the conditional factor demandcurves and the envelope theorem also implies that

C4(r, w, q, y) = λ∗(r, w, q, y),

so that the Lagrange multiplier λ∗(r, w, q, y) is again a measure of marginal cost.

20

Next, let’s use Shephard’s lemma to deduce another property of the conditional factordemand curves. Since

C1(r, w, q, y) = k∗(r, w, q, y),

it follows thatC12(r, w, q, y) = k∗2(r, w, q, y).

And sinceC2(r, w, q, y) = l∗(r, w, q, y),

it follows thatC21(r, w, q, y) = l∗1(r, w, q, y).

But, since the symmetry of the second partial derivatives of C implies

C12(r, w, q, y) = C21(r, w, q, y),

Shephard’s lemma can be viewed as having, as a corollary, the “reciprocity” conditionfor conditional factor demand curves:

k∗2(r, w, q, y) = l∗1(r, w, q, y),

or, equivalently,∂k∗(r, w, q, y)

∂w=∂l∗(r, w, q, y)

∂r.

Now consider a “short-run” version of the firm’s problem, treating the capital stock k asfixed at some pre-determined level k̄:

minl,m

rk̄ + wl + qm subject to f(k̄, l,m) ≥ y.

With the Lagrangian for the short-run problem defined as

L(l,m, µ) = rk̄ + wl + qm− µ[f(k̄, l,m)− y],

the first-order conditions are

w − µsf2(k̄, ls,ms) = 0

andq − µsf3(k̄, ls,ms) = 0

and the binding constraint is

f(k̄, ls,ms)− y = 0.

Interestingly, none of these optimality conditions depends on the rental rate r for capital.Hence, we can solve for short-run conditional factor demand curves of the form

ls = ls(w, q, y, k̄)

21

andms = ms(w, q, y, k̄),

and use another application of the envelope theorem, this time to the short-run prob-lem, to interpret

µs = µs(w, q, y, k̄)

as a measure of “short-run marginal cost.”

Le Chatelier’s principle says that labor demand should be more responsive to a change inthe wage rate w in the long run than in the short run since, intuitively, in the long runthe firm has the chance to substitute capital for labor in response to that change inthe wage.

To show the sense in which this conjecture proves true, suppose that the fixed, short-runvalue of k̄ just happens to equal the optimal value k∗ that would have been chosenanyway in the long run:

k̄ = k∗(r, w, q, y).

In this case, a comparison of the Kuhn-Tucker conditions for the short-run problemwith those for the long-run problem confirms that the firm will choose a value of ls inthe short run equal to the long-run value l∗:

l∗(r, w, q, y) = ls[w, q, y, k∗(r, w, q, y)].

Differentiate both sides of this expression with respect to w to obtain

l∗2(r, w, q, y) = ls1[w, q, y, k∗(r, w, q, y)] + ls4[w, q, y, k∗(r, w, y)]k∗2(r, w, q, y).

The term on the left-hand side of this expression measures the long-run responsivenessof labor demand to a change in the wage; the first term on the right-hand side ofthis expression measures the short-run responsiveness of labor demand to a change inthe wage. The second-order conditions for the long-run and short-run problems willgenerally imply that under “typical” conditions, both of these terms are negative: whenthe wage goes up, labor demand falls. But, is the long-run response “more negative?”This depends on the sign of the second term on the right-hand side, which in turndepends on ls4, measuring the response of short-run labor demand to a change in k̄,and k∗2, measuring the response of long-run capital demand to a change in w. The signsof both terms would seem to be ambiguous, dependent in some loose sense on whether,in the long run, the capital and labor inputs are “complements” or “substitutes.”

Yet, as it turns out, something more definite can be said. Return once more to

l∗(r, w, q, y) = ls[w, q, y, k∗(r, w, y)],

but this time differentiate both sides with respect to r to obtain

l∗1(r, w, q, y) = ls4[w, q, y, k∗(r, w, y)]k∗1(r, w, q, y).

22

Rearrange this expression so that it reads

ls4[w, q, y, k∗(r, w, q, y)] =l∗1(r, w, q, y)

k∗1(r, w, q, y),

and substitute this result into the previous one

l∗2(r, w, q, y) = ls1[w, q, y, k∗(r, w, q, y)] + ls4[w, q, y, k∗(r, w, q, y)]k∗2(r, w, q, y).

to get

l∗2(r, w, q, y) = ls1[w, q, y, k∗(r, w, q, y)] +l∗1(r, w, q, y)k∗2(r, w, q, y)

k∗1(r, w, q, y).

Now it is clear that the second term must be negative. This is because the expressionin the numerator is, by the reciprocity condition implied by Shephard’s lemma, equalto the perfect square [l∗1(r, w, q, y)]2; meanwhile, the term in the denominator willbe negative, as implied by the second-order conditions to the long-run problem: theoptimal capital stock falls when the rental rate for capital rises.

Thus, it must be thatl∗2(r, w, q, y) ≤ ls1[w, q, y, k∗(r, w, y)]

and, since both of these terms are negative,∣∣∣∣∂l∗(r, w, q, y)

∂w

∣∣∣∣ ≥ ∣∣∣∣∂ls(w, q, y, k∗)∂w

∣∣∣∣ ,which is the relationship we were after: labor demand is more responsive to a changein the wage in the long run than in the short run. Note, however, that this is a resultthat only holds locally, when k̄ is sufficiently close to k∗(r, w, q, y).

4 Generalizing the Basic Results

4.1 The Kuhn-Tucker Theorem

Our “simple” version of the Kuhn-Tucker theorem applies to a problem with only one choicevariable and one constraint.

Section 19.6 of Simon and Blume’s book develops a proof for the more general case, withn choice variables and m constraints. Their proof makes repeated, clever use of theimplicit function theorem, which makes the arguments surprisingly short but also worksto obscure some of the intuition provided by the analysis of the simplest case.

Nevertheless, having gained the intuition the intuition from working through the simplecase, it is useful to see how the result extends.

Simon and Blume (Chapter 15) and Acemoglu (Appendix A) both present fairly generalstatements of the implicit function theorem. The special case or application of theirresults that we will need works as follows.

23

Consider a system of n equations, involving n “endogenous” variables y1, y2, . . . , yn and n“exogenous” variables c1, c2, . . . cn:

H1(y1, y2, . . . , yn) = c1,

H2(y1, y2, . . . , yn) = c2,...

Hm(y1, y2, . . . , yn) = cn.

Now suppose that for a specific set of values c∗1, c∗2, . . . c

∗n for the exogenous variables, all

the equations in the system are satisfied with the endogenous variables set equal toy∗1, y

∗2, . . . , y

∗n, so that

H1(y∗1, y∗2, . . . , y

∗n) = c∗1,

H2(y∗1, y∗2, . . . , y

∗n) = c∗2,

...

Hn(y∗1, y∗2, . . . , y

∗n) = c∗n.

Assume that each function Hi, i = 1, . . . , n, is continuously differentiable and that the n×nmatrix of derivatives

∂H1/∂y1 · · · ∂H1/∂yn∂H2/∂y1 · · · ∂H2/∂yn

.... . .

...∂Hn/∂y1 · · · ∂Hn/∂yn

is nonsingular at y∗1, y

∗2, . . . , y

∗n.

Then there exist continuously differentiable functions

y1(c1, c2, . . . , cn),

y2(c1, c2, . . . , cn),...

yn(c1, c2 . . . , cn),

defined in an open subset C of Rn containing (c∗1, c∗2, . . . , c

∗n), such that

H1(y1(c1, c2, . . . , cn), y2(c1, c2, . . . , cn), . . . , yn(c1, c2, . . . , cn)) = c1,

H2(y1(c1, c2, . . . , cn), y2(c1, c2, . . . , cn), . . . , yn(c1, c2, . . . , cn)) = c2,...

Hn(y1(c1, c2, . . . , cn), y2(c1, c2, . . . , cn), . . . , yn(c1, c2, . . . , cn)) = cn.

for all (c1, c2, . . . , cn) ∈ C.

24

With this result in hand, consider the following generalized version of the Kuhn-Tuckertheorem we proved earlier. Let there be n choice variables, x1, x2, . . . , xn. The objectivefunction F : Rn → R is continuously differentiable, as are the m functions Gj : Rn →R, j = 1, 2, . . . ,m that enter into the constraints

cj ≥ Gj(x1, x2, . . . , xn),

where cj ∈ R for all j = 1, 2, . . . ,m.

The problem can be stated as:

maxx1,x2,...,xn

F (x1, x2, . . . , xn) subject to cj ≥ Gj(x1, x2, . . . , xn) for all j = 1, 2, . . . ,m.

Note that, typically, m ≤ n will have to hold so that there is a set of values for thechoice variables that satisfy all of the constraints.

To define the Lagrangian, introduce the multipliers λj, j = 1, 2, . . . ,m, one for each con-straint. Then

L(x1, x2, . . . , xn, λ1, λ2, . . . , λm) = F (x1, x2, . . . , xn) +m∑j=1

λj[cj −Gj(x1, x2, . . . , xn)].

Theorem (Kuhn-Tucker) Suppose that x∗1, x∗2, . . . , x

∗n maximize F (x1, x2, . . . , xn) sub-

ject to cj ≥ Gj(x1, x2, . . . , xn) for all j = 1, 2, . . . ,m, where F and the Gj’s are allcontinuously differentiable. Suppose (without loss of generality) that the first m̄ ≤ mconstraints bind at the optimum and that the remaining m − m̄ ≥ 0 constraints arenonbinding, and assume that the m̄× n matrix of derivatives

G1,1(x∗1, x∗2, . . . , x

∗n) . . . G1,n(x∗1, x

∗2, . . . , x

∗n)

G2,1(x∗1, x∗2, . . . , x

∗n) . . . G2,n(x∗1, x

∗2, . . . , x

∗n)

.... . .

...Gm̄,1(x∗1, x

∗2, . . . , x

∗n) . . . Gm̄,n(x∗1, x

∗2, . . . , x

∗n)

, (12)

where Gj,i = ∂Gj/∂xi, has rank m̄. Then there exist values λ∗1, λ∗2, . . . , λ

∗m that, to-

gether with x∗1, x∗2, . . . , x

∗n, satisfy:

Li(x∗1, x∗2, . . . , x

∗n, λ

∗1, λ∗2, . . . , λ

∗n) = Fi(x

∗1, x∗2, . . . , x

∗n)

−m∑j=1

λ∗jGj,i(x∗1, x∗2, . . . , x

∗n) = 0

(13)

for i = 1, 2, . . . , n,

Ln+j(x∗1, x∗2, . . . , x

∗n, λ

∗1, λ∗2, . . . , λ

∗n) = cj −Gj(x

∗1, x∗2, . . . , x

∗n) ≥ 0, (14)

for j = 1, 2, . . . ,m,λ∗j ≥ 0, (15)

for j = 1, 2, . . . ,m, and

λ∗j [cj −Gj(x∗1, x∗2, . . . , x

∗n)] = 0, (16)

for j = 1, 2, . . . ,m.

25

Proof To begin, set the multipliers λ∗m̄+1, λ∗m̄+2, . . . , λ

∗m associated with the nonbinding

contraints equal to zero. Since each of the functions Gj, j = m̄ + 1, m̄ + 2, . . . ,m, iscontinuously differentiable, sufficiently small adjustments in the choice variables canbe made without violating these m− m̄ constraints or causing any of them to becomebinding.

Next, note that the m̄+ 1× n matrixF1(x∗1, x

∗2, . . . , x

∗n) . . . Fn(x∗1, x

∗2, . . . , x

∗n)

G1,1(x∗1, x∗2, . . . , x

∗n) . . . G1,n(x∗1, x

∗2, . . . , x

∗n)

G2,1(x∗1, x∗2, . . . , x

∗n) . . . G2,n(x∗1, x

∗2, . . . , x

∗n)

.... . .

...Gm̄,1(x∗1, x

∗2, . . . , x

∗n) . . . Gm̄,n(x∗1, x

∗2, . . . , x

∗n)

. (17)

must have rank m̄ < m̄+ 1. To see why, consider the system of equations

F (x1, x2, . . . , xn) = y∗

G1(x1, x2, . . . , xn) = c1

G2(x1, x2, . . . , xn) = c2

...

Gm̄(x1, x2, . . . , xn) = cm̄.

With y∗ set equal to the maximized value of the objective function,

y∗ = F (x∗1, x∗2, . . . , x

∗n),

each of these m̄+ 1 equations holds when the functions are evaluated at x∗1, x∗2, . . . , x

∗n.

In this case, the implicit function theorem implies that it should be possible to adjustthe values of m̄+1 of the choice variables so to find a new set of values x∗∗1 , x

∗∗2 , . . . , x

∗∗n

such that

F (x∗∗1 , x∗∗2 , . . . , x

∗∗n ) = y∗ + ε

G1(x∗∗1 , x∗∗2 , . . . , x

∗∗n ) = c1

G2(x∗∗1 , x∗∗2 , . . . , x

∗∗n ) = c2

...

Gm̄(x∗∗1 , x∗∗2 , . . . , x

∗∗n ) = cm̄.

for a strictly positive but sufficiently small value of ε. But this contradicts the assump-tion that x∗1, x

∗2, . . . , x

∗n solves the constrained optimization problem.

Since the matrix in (17) has rank m̄ < m̄ + 1, its m̄ + 1 rows must be linearly dependent.

26

Hence, there exist scalars α0, α1, . . . αm̄, at least one of which is nonzero, such that0...0

= α0

F1(x∗1, x∗2, . . . , x

∗n)

...Fn(x∗1, x

∗2, . . . , x

∗n)

+ α1

G1,1(x∗1, x∗2, . . . , x

∗n)

...G1,n(x∗1, x

∗2, . . . , x

∗n)

+ . . .+ αm̄

Gm̄,1(x∗1, x∗2, . . . , x

∗n)

...Gm̄,n(x∗1, x

∗2, . . . , x

∗n)

.(18)

Moreover, in (18), α0 6= 0, since otherwise, the matrix in (12) would have rank lessthan m̄.

Thus, for j = 1, 2, . . . , m̄, set λ∗j = −αj/α0. With these settings for λ∗1, λ∗2, . . . , λ

∗m̄, plus the

settings λ∗m̄+1 = λ∗m̄+2 = λ∗m = 0 chosen earlier, (18) implies that (13) must hold forall i = 1, 2, . . . , n. Clearly, (14) and (16) are satisfied for all j = 1, 2, . . . ,m, and (15)holds for all j = m̄ + 1, m̄ + 2, . . . ,m. So it only remains to show that (15) holds forj = 1, 2, . . . , m̄.

To see that these last conditions must hold, consider the system of equations

G1(x1, x2, . . . , xn) = c1 − δG2(x1, x2, . . . , xn) = c2

...

Gm̄(x1, x2, . . . , xn) = cm̄,

(19)

where δ ≥ 0. These equations hold, with δ = 0, at x∗1, x∗2, . . . , x

∗n. And since the matrix

in (12) has rank m̄, the implicit function theorem implies that there are functionsx1(δ), x2(δ), . . . , xn(δ) such that the same equations hold for all sufficiently small valuesof δ.

Since c1 − δ ≤ c1, the choices x1(δ), x2(δ), . . . , xn(δ) satisfy all of the constraints fromthe original optimization problem. And since, by assumption, x1(0) = x∗1, x2(0) =x∗2, . . . , xn(0) = x∗n maximizes the objective function subject to the constraints, it mustbe that

dF (x1(δ), x2(δ), . . . , xn(δ))

dδ

∣∣∣∣δ=0

=n∑i=1

Fi(x∗1, x∗2, . . . , x

∗n)x′i(0) ≤ 0. (20)

In addition, the equations in (19) implicitly defining x1(δ), x2(δ), . . . , xn(δ) imply

dG1(x1(δ), x2(δ), . . . , xn(δ))

dδ

∣∣∣∣δ=0

=n∑i=1

G1,i(x∗1, x∗2, . . . , x

∗n)x′i(0) = −1 (21)

anddGj(x1(δ), x2(δ), . . . , xn(δ))

dδ

∣∣∣∣δ=0

=n∑i=1

Gj,i(x∗1, x∗2, . . . , x

∗n)x′i(0) = 0 (22)

for j = 2, 3, . . . , m̄.

27

Putting all these results together, (13) implies

0 = Fi(x∗1, x∗2, . . . , x

∗n)−

m∑j=1

λ∗jGj,i(x∗1, x∗2, . . . , x

∗n).

for all i = 1, 2, . . . , n. Multiplying each of these equations by x′i(0) and summing overall i yields

0 =n∑i=1

Fi(x∗1, x∗2, . . . , x

∗n)x′i(0)−

n∑i=1

m∑j=1

λ∗jGj,i(x∗1, x∗2, . . . , x

∗n)x′i(0),

or

0 =n∑i=1

Fi(x∗1, x∗2, . . . , x

∗n)x′i(0)−

m∑j=1

λ∗j

[n∑i=1

Gj,i(x∗1, x∗2, . . . , x

∗n)x′i(0)

],

or, since λ∗j = 0 for j = m̄+ 1, m̄+ 2, . . . ,m,

0 =n∑i=1

Fi(x∗1, x∗2, . . . , x

∗n)x′i(0)−

m̄∑j=1

λ∗j

[n∑i=1

Gj,i(x∗1, x∗2, . . . , x

∗n)x′i(0)

].

In light of (21) and (22), this last equation simplifies to

0 =n∑i=1

Fi(x∗1, x∗2, . . . , x

∗n)x′i(0) + λ∗1.

And hence, in light of (20),λ∗1 ≥ 0.

Analogous arguments show thatλ∗j ≥ 0

for j = 2, 3, . . . , m̄ as well, completing the proof.

4.2 The Envelope Theorem

Proving a generalized version of the envelope theorem requires no new ideas, just repeatedapplications of the previous ones.

Consider, again, the constrained optimization problem with n choice variables and m con-straints:

maxx1,x2,...,xn

F (x1, x2, . . . , xn) subject to cj ≥ Gj(x1, x2, . . . , xn) for all j = 1, 2, . . . ,m.

Now extend this problem by allowing the functions F and Gj, j = 1, 2, . . . ,m, to dependon a parameter θ ∈ R:

maxx1,x2,...,xn F (x1, x2, . . . , xn, θ) subject to

cj ≥ Gj(x1, x2, . . . , xn, θ) for all j = 1, 2, . . . ,m.

28

Just as before, define the maximum value function V : R→ R as

V (θ) = maxx1,x2,...,xn

F (x1, x2, . . . , xn, θ)

subject to cj ≥ Gj(x1, x2, . . . , xn, θ) for all j = 1, 2, . . . ,m.

Note that V is still a function of the single parameter θ, since the n choice variables are“optimized out.” Put another way, evaluating V requires the same two-step procedureas before:

First, given θ, find the values x∗1(θ), x∗2(θ), . . . , x∗n(θ) that solve the constrained opti-mization problem.

Second, substitute these values x∗1(θ), x∗2(θ), . . . , x∗n(θ), together with the given valueof θ, into the objective function to obtain

V (θ) = F (x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ).

And just as before, the envelope theorem tells us that we can calculate the derivative V ′(θ)of the maximum value function while ignoring the dependence of x∗1, x

∗2, . . . , x

∗n and

λ∗1, λ∗2, . . . , λ

∗m on θ, provided we invoke the complementary slackness conditions (16)

to add the sum of all of the multipliers times all of the constraints to the objectivefunction before differentiating through by θ.

Theorem (Envelope) Let F and Gj, j = 1, 2, . . . ,m, be continuously differentiable func-tions of x1, x2, . . . , xn and θ. For any value of θ, let x∗1(θ), x∗2(θ), . . . , x∗n(θ) maximizeF (x1, x2, . . . , xn, θ) subject to cj ≥ Gj(x1, x2, . . . , xn, θ) for all j = 1, 2, . . . ,m, and letλ∗1(θ), λ∗2(θ), . . . , λ∗m(θ) be the associated values of the Lagrange multipliers. Suppose,further, that x∗1(θ), x∗2(θ), . . . , x∗n(θ) and λ∗1(θ), λ∗2(θ), . . . , λ∗m(θ) are all continuously dif-ferentiable functions, and that the m̄(θ)×m matrix of derivatives

G1,1(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ) . . . G1,n(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)G2,1(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ) . . . G2,n(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)

.... . .

...Gm̄(θ),1(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ) . . . Gm̄(θ),n(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)

associated with the m̄(θ) ≤ m binding constraints has rank m̄(θ) for each value of θ.Then the maximum value function defined by

V (θ) = maxx1,x2,...,xn

F (x1, x2, . . . , xn, θ)

subject to cj ≥ Gj(x1, x2, . . . , xn, θ) for all j = 1, 2, . . . ,m

satisfies

V ′(θ) = Fn+1(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)

−m∑j=1

λ∗j(θ)Gj,n+1(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ).(23)

29

Proof The Kuhn-Tucker theorem implies that for any given value of θ,

Fi(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)−

m∑j=1

λ∗j(θ)Gj,i(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ) = 0 (13)

for i = 1, 2, . . . , n, and

λ∗j(θ)[cj −Gj(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)] = 0, (16)

for j = 1, 2, . . . ,m must hold.

In light of (16),

V (θ) = F (x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ) +m∑j=1

λ∗j(θ)[cj −Gj(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)].

Differentiating both sides of this expression by θ yields

V ′(θ) =n∑i=1

Fi(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)x∗′(θ)

+Fn+1(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)

+m∑j=1

λ∗′j (θ)[cj −Gj(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)]

−n∑i=1

m∑j=1

λ∗j(θ)Gj,i(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)x∗′(θ)

−m∑j=1

λ∗j(θ)Gj,n+1(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ).

which shows that, in principle, we must take the dependence of x∗1(θ), x∗2(θ), . . . , x∗n(θ)and λ∗1(θ), λ∗2(θ), . . . , λ∗m(θ) on θ into account when calculating V ′(θ).

Note, however, that (13) implies that the sums in the first and fourth lines of this lastexpression together equal zero. Hence, to show that (23) holds, it only remains toshow that

m∑j=1

λ∗′j (θ)[cj −Gj(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)] = 0

and this is true if

λ∗′j (θ)[cj −Gj(x∗1(θ), x∗2(θ), . . . , x∗n(θ), θ)] = 0 (24)

for all j = 1, 2, . . . ,m.

Clearly, (24) holds for any θ such that constraint j is binding.

30

For θ such that constraint j is not binding, (16) implies that λ∗j(θ) = 0. Furthermore, bythe continuity of Gj and xi(θ), i = 1, 2, . . . , n, if constraint j does not bind at θ, thereexists an ε∗ > 0 such that constraint j does not bind for all θ+ ε with ε∗ > |ε|. Hence,

λ∗′j (θ) = limε→0

λ∗j(θ + ε)− λ∗j(θ)ε

= limε→0

0

ε= 0,

and once again it becomes apparent that (24) must hold. Hence, (23) must hold aswell.

5 Berge’s Maximum Theorem

References:

Acemoglu, Appendix A.

Nancy L. Stokey and Robert E. Lucas, Jr. with Edward C. Prescott, RecursiveMethods in Economic Dynamics, 1989, Chapter 3.

Efe Ok, Real Analysis with Economic Applications, 2007, Sections E.1-E.3.

Claude Berge. Topological Spaces: Including a Treatment of Multi-Valued Functions,Vector Spaces, and Convexity, 1963.

Thus far, we have assumed that solutions to our constrained maximization problems existand vary smoothly with the parameters of interest. Claude Berge’s Maximum Theo-rem identifies a set of restrictions that can be imposed on the objective function andconstraint set for any such problem to guarantee that these results hold. Intuitively,we can view Berge’s theorem as a natural extension of Weierstrass’s extreme value the-orem: that any continuous function attains a maximum and a minimum on a compactset. Generalizing this result to apply to the smooth dependence of the solution on theparameters, however, requires a non-trivial initial “investment.” We will need to beginby specifying several notions of continuity for correspondences, that is, multi-valuedfunctions.

But let’s start by fixing the notation. Let x ∈ X ⊆ Rn be a vector of n choice variablesand let θ ∈ Θ ⊆ Rm be a vector of m parameters. Let the function F : X × Θ → Rdenote the objective function and let the correspondence G : Θ→ X describe, for anygiven θ ∈ Θ, the feasible set of values for x, denoted by G(θ) ⊆ X. We can now writethe optimization problem compactly as

supx∈G(θ)

F (x, θ).

Note that by casting the problem initially as finding the supremum and not the max-imum, we avoid presupposing that the maximum is actually attained: here, we wantthis as a result, not an assumption.

31

We must now deal directly with the fact that because G is not a single-valued, standardnotions of continuity, which apply to functions, will not apply. To simplify the analysis,it is helpful at the outset to assume that G is compact-valued, meaning that for allθ ∈ Θ, the set G(θ) ⊆ X ⊆ Rn is compact (or, equivalently, by the Heine-Boreltheorem, closed and bounded). Each of the continuity concepts introduced below,except one, applies even if G is not compact-valued. But, because this restriction isimposed by Berge’s theorem, we might as well make the assumption from the start.

The first notion, of upper hemi-continuity, can be cast equivalently in terms of sets orsequences as follows.

Definition The correspondence G is upper hemi-continuous at θ ∈ Θ if G(θ) is non-empty and if, for every open set X ′ ⊆ X with G(θ) ⊆ X ′ (i.e., any open subset of Xcontaining G(θ)), there exists a δ > 0 such that for every θ′ ∈ Nδ(θ), G(θ′) ⊆ X ′ (i.e.,a neighborhood of θ so that G(θ′) is also in X ′ for all θ′ in that neighborhood).

Definition The compact-valued correspondence G is upper hemi-continuous at θ ∈ Θ ifG(θ) is non-empty and if, for every sequence {θj} such that θj → θ and every sequence{xj} such that xj ∈ G(θj) for all j, there exists a convergent subsequence {xjk} suchthat xjk → x ∈ G(θ).

Note that only the second definition requires G to be compact-valued. If G is not compact-valued, the conditions imposed on the sequences in the second definition are sufficient,but not necessary, for G to be upper hemi-continuous as described in the first definition.

The second notion, lower hemi-continuity, can be also cast either in terms of sets or se-quences as follows.

Definition The correspondence G is lower hemi-continuous at θ ∈ Θ if G(θ) is non-emptyand if, for every open set X ′ ⊆ X with G(θ) ∩ X ′ 6= ∅ (i.e., any open subset of Xintersecting with G(θ)), there exists a δ > 0 such that for every θ′ ∈ Nδ(θ), G(θ′)∩X ′ 6=∅ (i.e., a neighborhood of θ so that G(θ′) also intersects with X ′ for all θ′ in thatneighborhood).

Definition The correspondence G is lower hemi-continuous at θ ∈ Θ if G(θ) is non-emptyand if, for every x ∈ G(θ) and every sequence {θj} such that θj → θ, there is a numberJ ≥ 1 and a sequence {xj} such that xj ∈ G(θj) for all j ≥ J and xj → x.

Using either definition in each case, G is upper hemi-continuous on Θ if G(θ) if it is upperhemi-continuous at every θ ∈ Θ and G is lower hemi-continuous on Θ if it is lowerhemi-continuous at every θ ∈ Θ. Finally, G is continuous if it is both upper and lowerhemi-continuous.

Definition The correspondenceG is continuous at θ ∈ Θ if it is both upper hemi-continuousand lower hemi-continuous at θ. The correspondence G is continuous on Θ if it iscontinuous at every θ ∈ Θ.

32

Before moving on, it is worthwhile to try to build up intuition as to the behavior that isruled out by the requirement that a correspondence be upper or lower hemi-continuous.Loosely speaking, an upper hemi-continuous correspondence cannot suddenly “ex-plode.” In the figure below, the correspondence G is lower but not upper hemi-continuous at θ0.

Meanwhile, a lower hemi-continuous correspondence cannot suddenly “implode.” In thefigure below, the correspondence G is upper but not lower hemi-continuous.

These previous examples suggest that if G bounds x from above and below by two contin-uous functions, then G will be continuous. And indeed we can prove that this true.

33

Specifically, let f : Θ → R and g : Θ → R be two continuous functions satisfying f(θ) ≤g(θ) for all θ ∈ Θ. Then define G : Θ→ R as

G(θ) = {x ∈ R | f(θ) ≤ x ≤ g(θ)}.

Note first that G is nonempty since, for example, g(θ) ∈ G(θ) for all θ ∈ Θ.

Next, let’s show that G is upper hemi-continuous. To do this, fix θ ∈ Θ and let {θj} be anysequence with θj → θ. Since G is nonempty, we can also find a sequence {xj} such that

xj ∈ G(θj) for all j. Since θj → θ, there exists a bounded set Θ̂ ⊆ Θ ⊆ R such that,

for some J ≥ 1, all of the θj for j ≥ J and θ will be contained in Θ̂. Moreover, thespecial structure of G in this case will imply that all of the xj for j ≥ J will also lie ina bounded subset of R. Thus, for all j ≥ J , all elements of the sequence {(θj, xj)} liein a bounded subset of R2 and, by the Bolzano-Weierstrass theorem, this sequence hasa convergent subsequence {(θjk , xjk)} with limit point (θ, x). And since each elementof this convergent sequence satisfies

f(θjk) ≤ xjk ≤ g(θjk),

the continuity of f and g guarantees that the limit point satisfies x ∈ G(θ) as well.

Finally, to show that G is lower hemi-continuous, fix θ ∈ Θ and x ∈ G(θ), and let {θj} beany sequence with θj → θ. If f(θ) = g(θ), then the sequence {xj} with xj = g(θj) forall j satisfies xj ∈ G(θj) for all j and, because g is continuous, xj → g(θ) = x as well.If, on the other hand, f(θ) < g(θ), note that

x− f(θ)

g(θ)− f(θ)

is well-defined and, since x ∈ G(θ), lies between zero and one. Therefore, the sequence{xj} with

xj =

[1− x− f(θ)

g(θ)− f(θ)

]f(θj) +

[x− f(θ)

g(θ)− f(θ)

]g(θj)

for all j satisfies xj ∈ G(θj) for all j and, since f and g are both continuous

xj →[1− x− f(θ)

g(θ)− f(θ)

]f(θ) +

[x− f(θ)

g(θ)− f(θ)

]g(θ) = f(θ) + x− f(θ) = x,

completing the proof.

Note that this last example also shows that if G(θ) is single-valued, with G(θ) = g(θ) forsome continuous function g, then G is both upper hemi-continuous and lower hemi-continuous. In fact, we can also show that the converse is true: if G(θ) is single-valued,with G(θ) = g(θ) for some function g, then g is continuous if G is either upper hemi-continuous or lower hemi-continuous.

To establish this, we will first show that if G is single-valued and upper hemi-continuous,then it must also be lower hemi-continuous. The proof is by contradiction. Supposethat G is single-valued, so that G(θ) = g(θ) for all θ ∈ Θ for some function g : Θ→ X.And suppose also G is upper hemi-continuous but not lower hemi-continuous.

34

In this case, fix θ ∈ Θ and {θj} with θj → θ. Because G is single-valued, {xj} is defineduniquely by xj = G(θj) = g(θj) for all j. But since G is not lower hemi-continuous,then xj cannot converge to x = G(θ) = g(θ). That is, there must exist an ε > 0 suchthat, for any J ≥ 1, there exists a j ≥ J such that |xj − x| ≥ ε. Using these values ofxj = g(θj), construct the subsequence {θjk} of {θj}; this subsequence has θjk → θ but,by construction, there is no subsequence of {xjk} converging to x. This shows thatG cannot be upper hemi-continuous: we have our contradiction, implying that if G issingle-valued and upper hemi-continuous, it must also be lower hemi-continuous.

Next, assume that G is single-valued, with G(θ) = g(θ) for all θ ∈ Θ for some functiong : Θ → X, and that G is lower hemi-continuous. We will show that g is continuous.Fix θ ∈ Θ and {θj} with θj → θ. Then x = G(θ) = g(θ) is uniquely-defined, as is thesequence {xj} with xj = G(θj) = g(θj) for all j. Since G is lower hemi-continuous,xj → x. This shows that for any sequence {θj} with θj → θ, the sequence {g(θj)}must have g(θj)→ g as required.

Thus, when G is single-valued, with G(θ) = g(θ) for all θ ∈ Θ, the correspondence G iscontinuous if and only if the function g is continuous, confirming that the notion of acontinuous correspondence generalizes, in a natural way, the more familiar definitionof a continuous function.

We can now state and prove Berge’s result, which requires F to be continuous in both xand θ and G to be compact-valued and continuous.

Theorem (Berge’s Maximum Theorem) Let X ⊆ Rn and Θ ⊆ Rm. Let F : X ×Θ→R be a continuous function, and let G : Θ → X be a compact-valued and continuouscorrespondence. The the maximum value function

V (θ) = maxx∈G(θ)

F (x, θ)

is well-defined and continuous, and the optimal policy correspondence

x∗(θ) = {x ∈ G(θ) | F (x, θ) = V (θ)}

is nonempty, compact-valued, and upper hemi-continuous.

Proof Fix θ ∈ Θ. Note first that since G(θ) is nonempty and compact, and since F (·, θ) iscontinuous, Weierstrass’s extreme value theorem implies that V (θ) is well-defined andthat x∗(θ) is nonempty.

Notice, next, that since x∗(θ) ⊆ G(θ) and G(θ) is compact, it follows that x∗(θ) is bounded.Let {xj} be a sequence with xj ∈ x∗(θ) for all j and xj → x. Since G(θ) is closed, itmust be that x ∈ G(θ). And since V (θ) = F (xj, θ) for all j and F is continuous, itfollows that F (x, θ) = V (θ). Hence, x ∈ x∗(θ), so x∗(θ) is closed. We now know thatx∗(θ) is nonempty and compact-valued for all θ ∈ Θ.

35

To see that x∗(θ) must be upper hemi-continuous, fix θ ∈ Θ and let {θj} be any sequencewith θj → θ. Choose xj ∈ x∗(θj) for all j. Since G is upper hemi-continuous, thereexists a subsequence {xjk} converging to x ∈ G(θ). Let x′ ∈ G(θ) as well. Since G islower hemi-continuous, there exists a sequence {x′jk} with x′jk ∈ G(θjk) and x′jk → x′.Since F (xjk , θjk) ≥ F (x′jk , θjk) for all k and F is continuous, it follows that F (x, θ) ≥F (x′, θ). And since this condition holds for any x′ ∈ G(θ), it follows that x ∈ x∗(θ), sothat x∗ is upper hemi-continuous.

Finally, to see that V is continuous, fix θ ∈ Θ, and let {θj} be any sequence with θj → θ.Choose xj ∈ x∗(θj) for all j. Let v̄ = lim supj V (θj) and v = lim infj V (θj). Thenthere exists a subsequence {xjk} such that v̄ = limk F (xjk , θjk). But since x∗ is upperhemi-continuous, these exists a subsequence of {xjkl} of {xjk} converging to x ∈ x∗(θ).Hence v̄ = liml F (xjkl , θjkl ) = F (x, θ) = V (θ). An analogous argument establishes thatv = V (θ). Hence V (θj)→ V (θ), completing the proof.

Note that if we assume, as well, that F (·, θ) is concave and G(θ) is convex for all θ ∈ Θ,then the optimal policy correspondence is x∗ is single-valued. Since Berge’s theoremimplies that, in general, x∗ is an upper hemi-continuous correspondence, in this caseour previous arguments show that x∗ is a continuous function as well.

As an application of the result, let’s consider the case of utility maximization with twogoods. The consumer takes income I and prices p1 and p2 as given, and solves

maxc1,c2

U(c1, c2) subject to I ≥ p1c1 + p2c2, c1 ≥ 0, c2 ≥ 0.

To apply the theorem, we need the utility function U to be continuous, and the corre-spondence

G(I, p1, p2) = {(c1, c2) ∈ R2 | I ≥ p1c1 + p2c2, c1 ≥ 0, c2 ≥ 0}

to be compact-valued and continuous. Clearly, for any (I, p1, p2) ∈ R3 with I > 0,p1 > 0 and p2 > 0, G is nonempty and compact-valued. So it only remains to confirmthat G is upper and lower hemi-continuous.

To see that G is upper hemi-continuous, fix the parameter values (I, p1, p2) with I > 0,p1 > 0, and p2 > 0 and let {(Ij, p1j, p2j)} be a sequence with Ij > 0, p1j > 0 and p2j > 0for all j with (Ij, p1j, p2j) → (I, p1, p2). Since G is non-empty, we can find a sequence{(c1j, c2j)} with (c1j, c2j) ∈ G(Ij, p1j, p2j) for all j. Now, since (Ij, p1j, p2j)→ (I, p1, p2),

we can also find a bounded set Θ̂ ⊆ Θ ⊆ R3, such that, for some J ≥ 1, all of the(Ij, p1j, p2j) for j ≥ J and (I, p1, p2) are contained in Θ̂. Note, too, that the specialstructure of G in this case implies that all of the (c1j, c2j) ∈ G(Ij, p1j, p2j) for j ≥ Jwill lie in a bounded subset of R2. Thus, for all j ≥ J , all elements of the sequence{(Ij, p1j, p2j, c1j, c2j)} lie in a bounded subset of R5 and, by the Bolzano-Weierstrasstheorem, this sequence has a convergent subsequence {(Ijk , p1jk , p2jk , c1jk , c2kk)} withlimit point (I, p1, p2, c1, c2). And since each element of this convergent subsequencesatisfies

Ijk ≥ p1jkc1jk + p2jkc2jk ,

36

c1jk ≥ 0, and c2jk ≥ 0, it is easy to see that the limit point will also have to satisfy(c1, c2) ∈ G(I, p1, p2).

To see that G is lower hemi-continuous as well, fix (I, p1, p2) with I > 0, p1 > 0, andp2 > 0 and (c1, c2) ∈ G(I, p1, p2), then let {(Ij, p1j, p2j)} be a sequence with Ij > 0,p1j > 0 and p2j > 0 for all j and (Ij, p1j, p2j) → (I, p1, p2). If c1 = c2 = 0, thensimply constructing the sequence {(c1j, c2j)} with c1j = c2j = 0 for all j will providethe needed example where (c1j, c2j) ∈ G(Ij, p1j, p2j) for all j and (c1j, c2j) → (c1, c2).If c1 and/or c2 is strictly positive, then construct the sequence {(c1j, c2j)} with

c1j =

(IjI

)(p1c1 + p2c2

p1jc1 + p2jc2

)c1

and

c2j =

(IjI

)(p1c1 + p2c2

p1jc1 + p2jc2

)c2

for all j. Note that c1j ≥ 0, c2j ≥ 0, and

p1jc1j + p2c2j =

(IjI

)(p1c1 + p2c2

p1jc1 + p2jc2

)(p1jc1 + p2jc2) = Ij

(p1c1 + p2c2

I

)≤ Ij,

so that (c1j, c2j) ∈ G(Ij, p1j, p2j) for all j. Moreover, (c1j, c2j) → (c1, c2), completingthe proof.

Thus, we can apply Berge’s theorem to the consumer’s problem and conclude that if theutility function is continuous:

The indirect utility function V (I, p1, p2) is well-defined and continuous.

And the demand correspondences c∗1(I, p1, p2) and c∗(I, p1, p2) are nonempty, compact-valued, and upper hemi-continuous.

And since the set G(I, p1, p2) is convex for all I > 0, p1 > 0, and p2, if we assume as well thatU is strictly concave, the Marshallian demand functions c∗1(I, p1, p2) and c∗(I, p1, p2)are well-defined (single-valued) and continuous.

37