19.

Constrained Optimization II

11/22/22

NB: Problems 2, 7, and 13 from Chapter 18 and problems 2 and 3 from

Chapter 19 are due on Tuesday, November 29.

We continue our investigation of constrained optimization, including

some the ideas surrounding the Kuhn-Tucker theory. We focus on four
main areas.
1. Dependence of the optimum and optimal values on model param-
eters, including the Envelope Theorem – page 3.
2. Second order conditions using bordered Hessians – page 19.
3. Smoothness of Optimal Functions – page 29.
4. Constraint qualification: Necessity and alternatives – page 35.
2 MATH METHODS

19.1 Envelope Theorems and Lagrange Multipliers

We’ve used multipliers to solve optimization problems, but we haven’t
stopped to ask: Does the multiplier mean anything? If so, what does the
multiplier mean?
That will eventually lead to questions about the relation between the
optimizers, the optimized value, and any parameters of the problem.
This is all summed up in the Envelope Theorem.1

1
Envelope theorems describe the relation between an optimized objective and various
parameters of the optimization problem. I don’t know their origin, but they have long
been used in economics. They can also be applied well beyond their traditional setting.
E.g., see Paul Milgrom and Ilya Segal (2002), Envelope theorems for arbitrary choice sets,
Econometrica 70, 583–601. Paul Samuelson had shown that Le Chatilier’s Principle
can be derived from an Envelope theorem. See Paul Samuelson (1947), Foundations
of Economic Analysis, Harvard Univ. Press, Cambridge, Mass.
 19. CONSTRAINED OPTIMIZATION II 3
19.1.1 Multiplier Interpretation
Let’s start with the multiplier question. We first examine it in a stripped
down model. Start with an objective function f(x, y) and a single con-
straint function h(x, y). That gives us the following maximization prob-
lem:
M(c) = max f(x, y)
(x,y)
(19.1.1)
s.t. h(x, y) = c

where M(c) = f x∗ (c), y∗(c) is the maximum value of f under the

constraint.
The optimality condition implicitly defines the multiplier µ.

Df = µDh.

Even in the simplest cases where it’s useful to use a multiplier, we must
have at least two goods and the first order condition is a covector equa-
tion. Let’s focus on the first component.

∂f ∂h ∂f ∂h
=µ or µ=
∂x ∂x ∂x ∂x

The units of µ are units of f per unit of h. This happens regardless of

which variable, x or y, is used to define µ. If we used y, we would get
the same units in the end.
That suggests the multiplier is the marginal value of whatever it is that is
being constrained, whatever h is. In this model we can also call it c. We
will calculate the marginal value of c and see if it is in fact the multiplier.
4 MATH METHODS
19.1.2 Finding the Multiplier: A Basic Envelope Theorem
We return to the maximization problem

M(c) = max f(x, y)

(x,y)
(19.1.1)
s.t. h(x, y) = c

and regard c is an adjustable parameter. The maximizer x∗ (c), y∗(c)

then depends
on c. The use of M(c) for the maximized value,

∗ ∗
f x (c), y (c) , emphasizes its dependence on the parameter c.
What we want to know is whether

′ dh ∗ ∗

i
µ = M (c) = f x (c), y (c) .
dc

In fact, it does.
Theorem 19.1.1. Let f and h be C1 functions on U ⊂ R2 . Let x∗ (c), y∗ (c)

be the solution to the maximization problem (19.1.1) and suppose x∗(c),

y∗ (c), and the multiplier µ∗(c) are continuously differentiable functions of
c. Suppose further that NDCQ holds at x∗(c), y∗ (c), µ∗(c) . Then

∗ ′ dh ∗ ∗

i
µ (c) = M (c) = f x (c), y (c) .
dc

Theorem 19.1.1 is a special case of the Envelope Theorem, which we

will encounter later.
 19. CONSTRAINED OPTIMIZATION II 5
19.1.3 Envelope Theorem Assumptions
It’s worth taking a moment to consider why the Envelope Theorem has
the assumptions it does.
The Set U. You might be curious why the Envelope Theorem involves
a subset U of R2 . Usually, we will be working with functions defined
on subsets of R2 . For example, both utility and production functions
are typically defined on R2+ . By using a subset U, we give the theorem
enough flexibility to handle such problems.
Constraint Qualification. Constraint qualification (NDCQ) must be satis-
fied so that we can use the Lagrangian to find the first order conditions.
See Theorems 18.11.1 18.18.1 and 18.35.1.
Differentiability. We require that the optimal choices and multipliers be
C1 functions as we will need to take their derivatives as part of the proof.
That leaves open the question of how we know these functions are C1,
which we will address later in this chapter.
6 MATH METHODS
19.1.4 A Basic Envelope Theorem
Theorem 19.1.1. Let f and h be C1 functions on U ⊂ R2 . Let x∗ (c), y∗ (c)

be the solution to the maximization problem (19.1.1) and suppose x∗(c),

y∗ (c), and the multiplier µ∗(c) are continuously differentiable

functions of
∗ ∗ ∗
c. Suppose further that NDCQ holds at x (c), y (c), µ (c) . Then

∗ ′ dh ∗ ∗

i
µ (c) = M (c) = f x (c), y (c) .
dc

Sketch of Proof. Since we will use the same method to prove several
variations on the basic Envelope Theorem, it’s worth summarizing our
method upfront.
Logically, the main part of the proof starts when we use the
Chain
Rule to calculate the derivative of the optimal value f x∗ (c), y∗(c) . We
then rewrite the equation twice: Once using the first order conditions
(19.1.2), and a second time using the results (19.1.3) from differentiating
the constraint. At that point the result pops out.
The formal proof is arranged somewhat differently. Since we know we
will need them, we pre-calculate both the first order conditions and the
derivative of the constraint. Once that is done, we are ready for the main
calculation, computing the derivative of the optimal value f x∗(c), y∗ (c) .

The pre-calculation allows the calculation to flow better, as we can plug

in our previous results when needed rather than interrupting the main
calculation to derive them.
 19. CONSTRAINED OPTIMIZATION II 7
19.1.5 Proof of Basic Envelope Theorem


Proof. The Lagrangian is L = f(x, y) − µ h(x, y) − c The first order
conditions are

∂L ∂f ∂h ∂L ∂f ∂h
0= = − µ∗ 0= = − µ∗
∂x ∂x ∂x ∂y ∂y ∂y
so
∂f ∂h ∂f ∂h
= µ∗ and = µ∗ (19.1.2)
∂x ∂x ∂y ∂y
with everything evaluated at (x,
y, µ) = x∗(c), y∗ (c), µ∗(c) . Differentiat-

ing the constraint h x∗(c), y∗ (c) = c, we obtain

∂h dx∗ ∂h dy∗
+ = 1. (19.1.3)
∂x dc ∂y dc

Applying the Chain Rule and using equations (19.1.2) and (19.1.3) we
find

′ dh ∗ ∗

i
M (c) = f x (c), y (c)
dc
∂f dx∗ ∂f dy∗
= +
∂x dc ∂y dc
  ∗   ∗
∂h dx ∂h dy
= µ∗ + µ∗ By (19.1.2)
∂x dc ∂y dc
∗
∂h dy∗
 
∗ ∂h dx
=µ +
∂x dc ∂y dc
= µ∗(c). By (19.1.3)
8 MATH METHODS
19.1.6 Marginal Utility of Income
Let’s apply Theorem 19.1.1 to the simple consumer’s problem we ex-
amined in the previous chapter.

v(p, m) = max u(x1, x2)

x
s.t. p1 x1 + p2 x2 = m.

We treat m as the parameter. Provided the solutions are differentiable,

(NDCQ holds automatically), we can apply Theorem 19.1.1 to find

∂v
(p, m) = µ∗ (m).
∂m

The multiplier belonging to the budget constraint is the marginal (indirect)

utility of income.
 19. CONSTRAINED OPTIMIZATION II 9
19.1.7 Marginal Cost
We can also consider a simplified version of the firm’s cost minimization
problem.

c(w, q) = min w· z
z
s.t. f(z) = q, .

We assume Df ≫ 0, ensuring that NDCQ holds. If the solutions are C1,

we have
∂c
(w, q) = µ∗ (q).
∂q
The multiplier here is the marginal cost.
Of course, we have to be careful drawing strong conclusions from
these simple versions of the utility maximization and cost minimization
problems. The details left out can be important. We will generalize
Theorem 19.1.1 to apply to the full consumer’s utility maximization
problem and firm’s cost minimization problem. That version can be used
to establish these results for the standard consumer’s utility maximization
problem and firm’s cost minimization problem.
10 MATH METHODS
19.1.8 Multiple Multipliers
Suppose we have many variables and multiple constraints, with multiple
multipliers. We consider the problem

M(c) = max f(x)

x
(19.1.4)
s.t. hi (x) = ci , for i = 1, . . . , ℓ

A version of Theorem 19.1.1 applies here
too. Each multiplier corre-

∗
sponds to a different derivative of f x (c) .
Theorem 19.1.2. Let U ⊂ Rm and suppose f : U → R and h : U → Rℓ
are C1 functions. Let x∗ (c) be the solution to the maximization problem
(19.1.4) and µ∗(c) the corresponding multipliers.
Suppose x∗ (c) and µ∗(c)
are C1 in c and that NDCQ holds at x∗(c), µ∗ (c) . Then
h
i
∗ ∗
µ (c) = Dc M(c) = Dc f x (c) .

Proof. The first order conditions are

Dx f = µ∗ Dx h (19.1.5)

where everything is
evaluated at the optimum. We also differentiate the

constraints h x∗(c) = c with respect to c, yielding

(Dx h)(Dc x∗) = Iℓ×ℓ (19.1.6)

Use the Chain Rule to differentiate f x∗(c), c and then substitute equa-

tions (19.1.5) and (19.1.6) to complete the proof:

h
i
∗
DcM(c) = Dc f x (c)
= (Dxf)(Dc x∗) Chain Rule
= µ∗(c)(Dxh)(Dc x∗ ) Equation (19.1.5)
= µ∗(c)Iℓ×ℓ Equation (19.1.6)
= µ∗(c)
 19. CONSTRAINED OPTIMIZATION II 11
19.1.9 Multipliers and Mixed Constraints
We move on to the general case of mixed constraints. This means we
have an optimization problem of this form.

M(b, c) = max f(x)

x
s.t. gi (x) ≤ bi , for i = 1, . . . , k (19.1.7)
hi (x) = ci , for i = 1, . . . , ℓ

It’s easy to see that if a constraint gh (x∗) ≤ bh does not bind at x∗, that
it will still not bind

if we change bh slightly. It follows that the maximized
∗
value, f x (b, c) , does not change either, and that its bh derivative must
be zero. Since constraint h does not bind, the corresponding multiplier
is zero by complementary slackness. The multiplier and derivative are
the same.
12 MATH METHODS
19.1.10 Theorem on Multipliers and Mixed Constraints
The proof of the following theorem is similar to the previous results.
BEWARE! The requirement that the maximized value and the maximiz-
ers be C1 can be difficult to meet when there are inequality constraints.
The problem is that you might be switching from one type of solution to
another when a constraint starts or finishes binding. The transition might
lead to a discontinuous derivative.
Theorem 19.1.3. Let U ⊂ Rm and suppose f : U → R, g : U → Rk , and
h : U → Rℓ are C1 functions. Let x∗ (b, c) be the solution to the max-
imization problem (19.1.7) and µ∗ (b, c) and λ∗ (b, c) the corresponding
multipliers. Suppose x∗ (b, c), µ∗ (b, c) and λ∗ (b, c) are C1 in (b, c).
Let ĝ be the vector of the k̂ binding constraints at x∗ and define
 ĝ(x) 
G(x) = .
h(x)

Suppose rank DG(x∗ ) = k̂ + ℓ holds (NDCQ).

Then
h
i
∗ ∗
λ (b, c) = Db M(b, c) = Db f x (b, c)
h
i
∗ ∗
µ (b, c) = Dc M(b, c) = Dc f x (b, c) .
 19. CONSTRAINED OPTIMIZATION II 13
19.1.11 Envelope Theorem
In the above theorem, we’ve been able to calculate the marginal value
of relaxing any of the constraints. However, we would like to calculate
marginal values for other parameters of the model, such as prices. For
that, we need a generalization known as the Envelope Theorem.
Let a ∈ Rn be a vector of parameters. Suppose the objective depends
on both x and a, with f(x, a) ∈ C1 and suppose we can also write
h(x, a) ∈ C1 where h(x, a) = 0 is the vector of equality constraints.
Let M(a) = f(x∗ (a), a) be the maximized value of f.
14 MATH METHODS
19.1.12 Redoing the Standard Calculations
The Lagrangian is
L = f(x, a) − µh(x, a)

We now employ the same procedure as before. The first order condi-
tions are
0 = Dx L = Dx f − µ∗ Dx h,

so
Dx f = µ∗ Dx h. (19.1.8)

We differentiate the constraint, which now has an extra term

h
i
∗
Da h x (a), a = (Dx h) (Da x∗) + Da h = 0

so
(Dx h) (Da x∗ ) = −Da h (19.1.9)

Then we wrap it all together using the Chain Rule and substitute equa-
tions (19.1.8) and (19.1.9)
h
i
∗
Da M(a) = Da f x (a), a Definition of M
= (Dxf) (Da x∗ ) + Da f Chain Rule
= µ∗ Dx h Da x∗ + Da f


Equation (19.1.8)
= −µ∗Da h + Da f Equation (19.1.9)
= Da f − µ∗ h
 
Rearrangement
= Da L x∗ (a), µ∗(a), a


Definition of L
 19. CONSTRAINED OPTIMIZATION II 15
19.1.13 Statement of Envelope Theorem
We now state an Envelope Theorem based on the above calculations.
We just need to add appropriate hypotheses so that the preceding cal-
culations are valid. In particular, the hypotheses must guarantee that our
first order conditions are correct and that the maximizers are C1 .
Envelope Theorem. Consider the maximization problem

M(a) = max f(x, a)

x
(19.1.10)
s.t. hj (x, a) = 0, for j = 1, . . . , ℓ

Let U ⊂ Rm and V ⊂ Rn and suppose f : U×V → R and h : U×V → Rℓ

are C1 functions. Let x∗(a) solve the maximization problem (19.1.10) for
any a ∈ V. Define the Lagrangian

L = f(x, a) − µh(x, a).

Suppose the maximizer x∗(a) and the corresponding Lagrange multipliers

µ∗ (a) are C1 functions of a, and that NDCQ holds. Then

Da M(a) = Da L x∗ (a), µ∗(a), a .

Theorem 19.1.2 is just the special case with h(x, a) = h(x) − a.

16 MATH METHODS
19.1.14 Indirect Utility and the Envelope Theorem
Consider a simple consumer’s problem.
◮ Example 19.1.4: Roy’s Identity. The indirect utility function v(p, m) is
defined by

v(p, m) = max u(x)

x
s.t. p·x = m

where u : Rm+ → R is a utility function, the prices obey p ≫ 0 and

income is positive, m > 0.
We set a = p. Then

∂v ∂u
− µ∗(a) xi (p, m) = −µ∗(a) xi(p, m).


= Dp L i =
∂pi ∂pi

We can use the fact that ∂v/∂m = µ∗ to establish Roy’s Identity:2

∂v/∂pi
xi (p, m) = − .
∂v/∂m

2
Roy’s identity was established by the French economist and econometrican René
Roy (1894–1977). See René Roy (1947), La distribution du revenu entre les divers
biens, Econometrica 15, 205–225.
 19. CONSTRAINED OPTIMIZATION II 17
19.1.15 Cost and the Envelope Theorem
Another example involves the cost function.3
◮ Example 19.1.5: Shephard’s Lemma. The cost function c(w, q) is de-
fined by

c(w, q) = min w· z
z
s.t. f(z) = q, .

where f : Rm+ → R is a production function and w ≫ 0 is the vector of

factor prices.
Since w doesn’t appear in the constraints, the Envelope Theorem tells
us that the derivative of the cost function is

Dw c(w, q) = Dw L = Dw (w·z) = z(w, q)T .

The row vector z(w, q)T is the vector of conditional factor demands. This
result is known as Shephard’s Lemma. ◭

3
Shephard’s Lemma is due to Ronald W. Shephard (1953), Cost and Production Func-
tions, Princeton University Press, Princeton, NJ. Cost functions had been used prior to
Shephard, but he may have been the first to use the modern definition. He used a
duality method to prove Shephard’s Lemma, based on the distance function.
Ronald W. Shephard (1912–1982) was an American economist who’s research con-
centrated on cost and production. He pioneered the use of duality based on the
distance function. In contrast, the usual economic duality is more directly related to
the conjugate function.
18 MATH METHODS
19.1.16 Expenditure and the Envelope Theorem
We can also apply this to the very similar expenditure function.4
◮ Example 19.1.6: Shephard-McKenzie Lemma. The expenditure func-
tion solves

e(p, ū) = min p·x

x
s.t. u(x) = ū, .

By the Envelope Theorem,

Dp e(p, ū) = Dp L = Dp (p·x) = h(p, ū)T ,

yielding the Hicksian or compensated demands h(p, ū). This result is

known as the Shephard-McKenzie Lemma. ◭

4
The expenditure function was introduced by Lionel McKenzie (1957), Demand
theory without a utility index, Rev. Econ. Studies 24, 185–189.
Lionel McKenzie (1919–2010) was an American economist known for proving the
existence of general equilibrium, for his work on demand theory, and for optimal growth
and turnpike theory.
 19. CONSTRAINED OPTIMIZATION II 19

19.2 Second Order Necessary and Sufficient Conditions

The problem with merely looking at the first order conditions is that they
are not enough to ensure that we have a maximum or minimum. They
are necessary, but not sufficient.
As we will see, there are second order conditions for constrained prob-
lems that are sufficient, but not necessary. As far as necessity is concerned,
the best we can do is find conditions that are locally necessary.
19.2.1 Constrained Negative Definite Lagrangians
As in the unconstrained case, we can use the second derivative to show
that a critical point of (19.1.4) is maximal. For the sufficient conditions,
we need to show that D2x L is negative definite, at least for those vectors
that are in the tangent space of the constraint set.
We can apply the bordered Hessian test to see if D2x L is negative
definite on the tangent space, Tx∗ M = ker Dx∗ h. Define the bordered
Hessian of the Lagrangian by

0ℓ×ℓ Dh(x∗ )
 
B= .
Dh(x∗ )T D2x L

where 0ℓ×ℓ is the ℓ × ℓ zero matrix.

20 MATH METHODS
19.2.2 Second Order Sufficient Conditions: Maximum
We can now state second order sufficient conditions for a constrained
strict local maximum.
Theorem 19.2.1. Let U ⊂ Rm and suppose f : U → R and h : U → Rℓ
are C2 functions. Form the Lagrangian


L = f(x) − µ h(x) − c .

Suppose that the following conditions are satisfied:

1. x∗ ∈ M = {x : h(x) = c}.
2. Dx L = 0 and Dµ L = 0 at (x∗, µ∗ ) (critical point).
3. rank Dh(x∗ ) = ℓ holds (NDCQ).
4. The Hessian of the Lagrangian with respect to x at (x∗, µ∗ ), D2x L(x∗, µ∗ ),
is negative

definite on the tangent space Tx∗ M = ker Dx∗ h = {v :
Dx∗ h v = 0}.
Assumptions (1) and (2) tell us that (x∗, µ∗ ) is a critical point of the problem
(19.1.4).
Then x∗ is a strict local maximum of f on the constraint set M.
Moreover, the conclusion remains true if (4) is replaced by (4’).
4’. The last (m − ℓ) leading principal minors of the bordered Hessian B
alternate in sign and the determinant of B obeys (−1)ℓ+m det B > 0.
 19. CONSTRAINED OPTIMIZATION II 21
19.2.3 The Natural Bordered Hessian
It is arguably more natural to use D2(µ,x)L rather than the bordered Hes-
sian. Here
∗
0 −Dh(x )
 
ℓ×ℓ
D2(µ,x)L(x∗, µ∗ ) =
−Dh(x∗ )T D2x L(x∗, µ∗ )

The extra minus signs do not affect whether the matrix D2x L is constrained
positive definite or constrained negative definite. Compared to the usual
bordered Hessian, the principal minors are unchanged because for each
column that is multiplied by (−1), the corresponding row is also multi-
plied by (−1). Together, they multiply each principal minor by (−1)2,
leaving every principal minor unchanged.
The bordered Hessian that appears in the proof of Theorem 19.2.1 is
the natural bordered Hessian, D2(µ,x)L, not our usual bordered Hessian.
Since they are equivalent in the sense that they both are negative definite
on the same subspaces and yield the same conditions on minors, it
doesn’t really affect the statement of the theorem.
22 MATH METHODS
19.2.4 Proof of Theorem 19.2.1
Proof. We start by noting that (4) and (4’) are equivalent. It doesn’t
matter which we use.
The rest of the proof proceeds by contradiction. Suppose that x∗ is
not a strict local maximum. Then for n = 1, 2, . . . , we can find an
xn 6= x∗ obeying h(xn ) = c with kxn − x∗k < 1/n and f(xn ) ≥ f(x∗ ).
Next define a sequence of unit vectors by

xj − x∗
uj = .
kxj − x∗ k

These all lie on the unit sphere Sm−1 = {x ∈ Rm : kxk = 1}, which is a
compact set. It follows that there is a convergent subsequence ujk . We
call its limit u∗ .
Since h is C2 , its first order Taylor polynomial with remainder about x∗
can be written5

h(xjk ) = h(x∗ ) + Dh(x∗)(xjk − x∗) + R1(xkj ; x∗) (19.2.11)

where R1(xkj ; x∗ ) = o kxkj − x∗k .

Using the fact h(xjk ) = c = h(x∗ ), we can simplify equation (19.2.11),

and then divide by kxjk − x∗k, obtaining

xj k − x∗ R1(xkj ; x∗ )
 
∗
0 = Dh(x ) + .
kxjk − x∗k kxjk − x∗k

Taking the limit as k → ∞ yields

0 = Dh(x∗ )u∗

showing that u∗ ∈ Tx∗ M.

Proof continues . . .

5
See section 30.29.
 19. CONSTRAINED OPTIMIZATION II 23
19.2.5 Remainder of Proof of Theorem 19.2.1
Remainder of Proof. Take the second order Taylor polynomial of the
Lagrangian L about x∗ with remainder S2.

L(xjk , µ∗ ) = L(x∗, µ∗ ) + DxL(x∗, µ∗ )(xjk − x∗ )

1
+ (xjk − x∗)T D2x L(x∗, µ∗) (xjk − x∗) + S2(xjk ; x∗ )
 
2

where S2(xjk − x∗ ) = o kxjk − x∗ k2 .

By the first order condition, Dx L(x∗, µ∗ ) = 0, so

1
L(xjk , µ∗ ) − L(x∗, µ∗ ) = (xjk − x∗ )T D2xL(x∗, µ∗ ) (xjk − x∗) + S2(xjk ; x∗).
 
2

Using the facts that h(xjk ) = c = h(x∗ ) for all k = 1, 2, 3 . . . , we

find L(xjk , µ∗ ) = f(xjk ) and L(x∗, µ∗ ) = f(x∗). By the contradiction
hypothesis,

0 ≤ f(xjk ) − f(x∗)
≤ L(xjk , µ∗) − L(x∗, µ∗ )
1
= (xjk − x∗ )T D2xL(x∗, µ∗ ) (xjk − x∗ ) + S2 (xjk ; x∗).
 
2

Dividing by kxjk − x∗ k2 and letting k → ∞, we find

1
0 ≤ u∗ T D2xL(x∗, µ∗ ) u∗ .
 
2

But u∗ is a unit vector with u∗ ∈ Tx∗ M, and D2xL(x∗, µ∗ ) is negative defi-

nite on such vectors by hypothesis. The right-hand side of the equation
is negative, so 0 < 0, which is impossible. This contradiction shows
that x∗ is a strict local max.
24 MATH METHODS
19.2.6 Second Order Sufficient Conditions: Minimum
Let (x∗, µ∗ ) be a critical point of the minimization problem analogous to
(19.1.4). If the Hessian is positive definite on Tx∗ M = ker Dx∗ h at a
critical point x∗ , then x∗ is a minimum.
Theorem 19.2.2. Let U ⊂ Rm and suppose f : U → R and h : U → Rℓ
are C2 functions. Form the Lagrangian


L = f(x) − µ c − h(x) .

Suppose that the following conditions are satisfied:

1. x∗ ∈ M = {x : h(x) = c}.
2. Dx L = 0 and Dµ L = 0 at (x∗, µ∗ ) (critical point).
3. rank Dh(x∗ ) = ℓ holds (NDCQ).
4. The Hessian of the Lagrangian with respect to x at (x∗, µ∗ ), D2x L(x∗, µ∗ )
is positive

definite on the tangent space Tx M = ker Dx h = {v :
∗ ∗

Dx∗ h v = 0}.
Assumptions (1) and (2) tell us that (x∗, µ∗ ) is a critical point of the mini-
mization problem analogous to (19.1.4).
Then x∗ is a strict local constrained minimum of f on M.
Moreover, the conclusion remains true if (4) is replaced by (4’).
4’. The last (m − ℓ) leading principal minors of the bordered Hessian B
have the same sign and the determinant of B obeys (−1)ℓ det B > 0.
Proof. Mimic the proof of Theorem 19.2.1.
 19. CONSTRAINED OPTIMIZATION II 25
19.2.7 2nd Order Conditions: Mixed Constraint Maxima
We’ll jump directly to the full mixed constraint version of the theorem.
Theorem 19.2.3. Let U ⊂ Rm and suppose the functions f : U → R,
g : U → Rk , and h : U → Rℓ are C2 functions. Suppose further that there
are k̂ binding constraints at x∗ . Let ĝ be the vector of binding inequality
constraints at x∗ with corresponding constants b̂. Set
 ĝ(x) 
G(x) =
h(x)

Suppose rank DG(x∗ ) = k̂ + ℓ holds (NDCQ) and define the constraint

set by M = {x : g(x) ≤ b, h(x) = c}. Form the Lagrangian:



L = f(x) − λ ĝ(x) − b̂ − µ h(x) − c .

Now suppose
1. x∗ ∈ M.
2. There are λ∗ ∈ Rk and µ∗ ∈ Rℓ with

Dx L = 0 at (x∗, λ∗ , µ∗ ),
λ∗ ≥ 0,
λ∗1 g1(x∗ ) − b1 ) = 0, . . . , λ∗k gk (x∗) − bk = 0,

h(x∗ ) = c.

3. The Hessian of the Lagrangian with respect to x at (x∗, λ∗ , µ∗ ),

D2x L(x∗, λ∗, µ∗ ) is negative

definite on the tangent space Tx M =
∗

ker Dx∗ G = {v : Dx∗ G v = 0}.

Assumptions (1) and (2) tell us (x∗, µ∗ ) is a critical point of the maximization
problem (19.1.4).
Then x∗ is a strict local constrained maximum of f on M.
Moreover, the conclusion remains true if (3) is replaced by (3’).
3’. The last (m − k̂ − ℓ) leading principal minors of the bordered Hessian
B alternate in sign and the determinant of B obeys (−1)m det B > 0.
26 MATH METHODS
19.2.8 2nd Order Conditions: Mixed Constraint Minima
The version for minima is only slightly different. Although the differences
are minor, they make all the difference in the world, resulting in minima
rather than maxima.
Theorem 19.2.4. Let U ⊂ Rm and suppose the functions f : U → R,
g : U → Rk , and h : U → Rℓ are C2 functions. Suppose further that there
are k̂ binding constraints at x∗ . Let ĝ be the vector of binding inequality
constraints at x∗ with corresponding constants b̂. Set
 ĝ(x) 
G(x) =
h(x)

Suppose rank DG(x∗ ) = k̂ + ℓ holds (NDCQ) and define the constraint

set by M = {x : g(x) ≥ b, h(x) = c}. Form the Lagrangian:



L = f(x) − λ ĝ(x) − b − µ h(x) − c .

Now suppose
1. x∗ ∈ M.
2. There are λ∗ ∈ Rk and µ∗ ∈ Rℓ with

Dx L = 0 at (x∗, λ∗ , µ∗ ),
λ∗ ≥ 0,
λ∗1 g1(x∗ ) − b1 ) = 0, . . . , λ∗k gk (x∗) − bk = 0,

h(x∗ ) = c.

3. The Hessian of the Lagrangian with respect to x at (x∗, λ∗ , µ∗ ),

D2x L(x∗, λ∗, µ∗ ) is positive

definite on the tangent space Tx M =
∗

ker Dx∗ G = {v : Dx∗ G v = 0}.

Then x∗ is a strict local constrained minimum of f on M.
Moreover, the conclusion remains true if (3) is replaced by (3’).
3’. The last (m − k̂ − ℓ) leading principal minors of the bordered
Hessian B have the same sign and the determinant of B obeys
(−1)k̂+ℓ det B > 0.
 19. CONSTRAINED OPTIMIZATION II 27
19.2.9 Second Order Necessary Conditions, Preliminaries
We will limit ourselves to a single theorem giving necessary conditions for
a local maximum in the mixed case. The local minimum case is similar.
This involves semidefinite matrices. There are also sufficient conditions
using semidefinite matrices, analogous to Theorem 17.8.1 and Theorem
17.8.1 Restated.
Let k̂ be the number of binding inequality constraints at x∗ and ĝ
the corresponding vector of binding inequality constraints. Form the
k̂ + ℓ-vector of binding constraints at x∗ ,
 ĝ(x) 
G(x) =
h(x)

Consider the bordered Hessian of the Lagrangian:

0
(k̂+ℓ)×(k̂+ℓ) DG(x∗) 
B= ∗ T
DG(x ) D2x∗ L(x∗, λ∗ , µ∗ )

Bordered Principal Submatrices. Let B̂r denote any r × r principal sub-

matrix with r > k̂ + ℓ, containing the 0(k̂+ℓ)×(k̂+ℓ) block in upper left hand
corner of B. We will refer to these as bordered principal submatrices of
rank r.
28 MATH METHODS
19.2.10 Second Order Necessary Conditions
We can now state the theorem.
Theorem 19.2.5. Let U ⊂ Rm and suppose the functions f : U → R,
g : U → Rk , and h : U → Rℓ are C2 functions. Suppose further that
there are k̂ binding inequality constraints at x∗. Let ĝ be the vector of
binding inequality constraint functions at x∗ with corresponding constants
b̂. Then set  ĝ(x) 
G(x) = ,
h(x)
which includes all binding constraints at x∗ .
Suppose that rank DG(x∗) = k̂ + ℓ holds (NDCQ) and define the con-
straint set by M = {x : g(x) ≤ b, h(x) = c}. Suppose that x∗ ∈ U0
is a local maximum of f on M, that (x∗, λ∗ , µ∗ ) is a critical point for the
Lagrangian


L = f(x) − λ ĝ(x) − b̂ − µ h(x) − c ,

and that the complementary slackness and non-negativity conditions hold.

Then the Hessian of the Lagrangian at (x∗ , λ∗ , µ∗ ) is negative semidefinite
on the tangent space Tx∗ = ker DG(x∗ ) = {v ∈ Rh+ℓ : DG(x∗) v = 0}.

Equivalently, either the minors det B̂r of any given size r > m−(k̂+ℓ) are
either all non-positive, or all non-negative. The generalized sign alternates
with r, and (−1)k̂+ℓ det B̂ > 0.
Simon and Blume punted on the determinant conditions. The best
thing to do is to look at the source. Our determinant condition is a
rewritten version of the one in Debreu (1952).6
Minimum. If x∗ is a minimum, we have to reverse the inequalities for g,
and adjust the Lagrangian accordingly. In that case, we can conclude
that the Hessian of the Lagrangian at (x∗ , λ∗ , µ∗ ) is positive semidefinite
on the tangent space Tx∗ .
Equivalently, the minors det B̂r with r > m − (k̂ + ℓ) are either all
non-positive, or all non-negative. and (−1)m−k̂−ℓ det B̂ > 0.

6
Debreu, Gerard (1952) Definite and semidefinite quadratic forms, Econometrica, 20,
295–300.
 19. CONSTRAINED OPTIMIZATION II 29

19.3 Smoothness and Optimization

Earlier, when considering the Envelope Theorem and related results, we
required that the maximum or minimum be a C1 function of various
parameters. One weakness of that approach is that we gave no method
of ensuring that the functions were C1 .
We are ready to remedy that and will give conditions on the primitives
of the model, the objectives and constraints, that guarantee the solutions
are C1 . The Implicit Function Theorem is the key tool for this.
19.3.1 Parameters and Optima
We start with an unconstrained problem to explore the basic ideas.
Consider a simple model with a single parameter a, and two variables,
(x, y) ∈ U ⊂ R2 . We consider the maximization problem

F(a) = max f(x, y, a).

(x,y)∈U

If U is compact, the Weierstrass Theorem guarantees a solution. If we

have an interior maximum, we know that the first order conditions must
be satisfied,
∂f ∂f
= 0 and = 0.
∂x ∂y
These equations may implicitly define functions x∗ (a) and y∗ (a) that
maximize f for at any given a.
30 MATH METHODS
19.3.2 What is Required for Smoothness?
For this to work, the Implicit Function Theorem requires that the (x, y)-
derivative of this vector equation be non-singular at x∗ (a), y∗(a), a .

That is, the Hessian matrix

D2(x,y)f x(a), y(a), a

must be non-singular.
If the second order sufficient conditions for a strict local maximum
are satisfied, the Hessian D2(x,y)f will be negative definite, and hence
non-singular.
So let’s write a theorem summing up what we proved. The Implicit
Function Theorem requires that our implicit function be defined by a
function that is itself C1. It also allows multiple exogenous variables, like
our variable a.
Differentiability is a local property, so the fact that the Implicit Function
Theorem only gives local results should not be a problem.
If we consider a similar minimization problem, the only real difference
is that the similar second order sufficient condition is that D2x f be positive
definite at x(a), which again means D2x f is non-singular.
Let’s add these considerations to our theorem.
 19. CONSTRAINED OPTIMIZATION II 31
19.3.3 Smoothness of Unconstrained Optima
Here’s a version of such a theorem, stated for x ∈ Rm .
Theorem 19.3.1. Let U ⊂ Rm and V ⊂ R. Suppose f : U × V → R is C2
in x ∈ U and C1 in a ∈ V. Suppose also there are (x0, a0) ∈ U0 × V 0
with Dx f(x0 , a0) = 0.
If the Hessian D2x f(x0, a0 ) is either positive definite or negative definite,
then there is an ε > 0 and a C1 function x∗ : Bε (a0 ) → Rm with x∗(a)
obeying

Dx f x(a), a = 0

on Bε (a0) with x(a0) = x0.

The function maximizes f if the Hessian is negative definite and mini-
mizes f if the Hessian is positive definite.
Proof. Left as an exercise to the reader. All of the key components
were mentioned on the previous page or included in previous sufficient
conditions.
32 MATH METHODS
19.3.4 Smoothness of Constrained Optima
It’s not much harder to show the solution to optimization problems with
equality constraints are smooth.
Consider the problem with ℓ equality constraints and one parameter,
a.
max f(x, a)
x
(19.3.12)
s.t h(x, a) = 0

where f : U → R and h : V × W → Rℓ with U, V ⊂ Rm and W ⊂ R.

If the NDCQ holds, rank Dxh = ℓ at the optimal x∗(a).
The Lagrangian is

L(x, µ, a) = f(x, a) − µh(x, a)

and any constrained optimum, x∗ (a) satisfies the first order conditions

DxL(x, µ, a) = 0, Dµ L(x, µ, a) = 0.

This is a system of m + ℓ equations (m from the first group, ℓ from the

second). There are also m + ℓ endogenous variables x and µ, along
with one exogenous variable a. Provided we know the problem has a
solution (e.g., by Weierstrass’s Theorem) the Implicit Function Theorem
tells us the solution is C1 if the derivative of this system with respect to
the endogenous variables is non-singular.
 19. CONSTRAINED OPTIMIZATION II 33
19.3.5 Implicit Functions and Constraint Qualification
We compute
0ℓ×ℓ −Dh
 

D2(µ,x)L =  . (19.3.13)
−DhT D2x L
It is the version of the bordered Hessian we previously encountered
following Theorem 19.2.1. This matrix will be singular if NDCQ fails.
That means that if we require D2(µ,x)L be invertible, we are also requiring
that the NDCQ is satisfied, as shown by the following theorem.
Theorem 19.3.2. Suppose D2(µ,x)L(x∗ ) given by equation (19.3.13) is in-
vertible. Then rank Dh(x∗ ) = ℓ.
Proof. Consider
∗


iu  − Dh(x ) v
h 
D2(µ,x)L(x∗) =
T
v − Dh(x∗ ) u + D2x L(x∗) v
 

for any u ∈ Rk and v ∈ Rm . If D2(µ,x)L is invertible, ℓ+m ≤ rank Dh(x∗ )+

m ≤ ℓ + m, so we must have rank Dh(x∗ ) = ℓ.
34 MATH METHODS
19.3.6 Smooth Parameterization Theorem
The above reasoning leads to the following theorem.
Theorem 19.3.3. Let x∗ (a) be the solution to the parameterized con-
strained maximization problem in equation (19.3.12) and let µ∗ (a) be
the corresponding vector of Lagrange multipliers. Fix a0. If
the bordered
Hessian matrix D2(µ,x)L is non-singular at x∗ (a0), µ∗ (a0), a0 , then
(a) the NDCQ holds at x∗(a0 ), µ∗ (a0), a0 , and

(b) x∗(a) and µ∗ (a) are C1 at a = a0.

This theorem can also be generalized to a vector of parameters. Cases
with inequality constraints are trickier. For the latter, it’s necessary to
focus on the binding constraints. Rather than considering a collection of
variations on a theorem, we return to the issue of constraint qualification.
 19. CONSTRAINED OPTIMIZATION II 35

19.4 Constraint Qualification

The example that began in section 18.33 showed us that if constraint
qualification fails, we may not be able to find the optimum.
In such cases we can try alternative methods of constraint qualification.
There’s even a theorem due to Fritz John where we can replace constraint
qualification with an additional multiplier for the objective.
36 MATH METHODS
19.4.1 A Multiplier for the Objective
One alternative is to use an additional multiplier for the objective.
Theorem 19.4.1. Let f and h be C1 functions on U ⊂ R2 . Suppose that
x∗ solves

max f(x)
x
s.t h(x) = c

Form the Lagrangian

L(x, µ0, µ1 ) = µ0 f(x) − µ1 h(x) − c ,

by including a multiplier for the objective function f. Then there exist

multipliers µ∗0 and µ∗1 such that
(a) µ∗0 is either 0 or 1,
(b) µ∗0 and µ∗1 are not both zero,
(c) (x∗, µ∗ ) satisfies the first order conditions

0 = Dx L = µ 0 D x f − µ 1 D x h
∂L
0= = c − h(x).
∂µ1

Proof. If Dx h 6= 0, constraint qualification (NDCQ) is satisfied, and

there is a solution with µ∗0 = 1 by Theorem 18.11.1.
If Dx h = 0, then set µ∗0 = 0.
In this case, the extra multiplier µ0 serves as a device to remind us to
check the critical points of the constraints as well as the objective.
 19. CONSTRAINED OPTIMIZATION II 37
19.4.2 Example 18.33 Revisited I
Recall that the problem

max x
(x,y)

s.t. x3 + y2 = 0.

showed us that without NDCQ constraint qualification, the standard first

order conditions might not be necessary for an optimum.
Let’s try Theorem 19.4.1 on it. Set up the modified Lagrangian

L = µ0 x − µ1(x3 + y2 ).

The first order conditions based on the modified Lagrangian are

µ0 = 3µ1x2
0 = −2µ1y
0 = x3 + y2.

We saw in section 18.33 that when µ0 = 1, there is no solution. What

if µ0 = 0?. Then the first of the first order condition becomes

0 = 3µ1 x2,

so either x = 0 or µ1 = 0.
38 MATH METHODS
19.4.3 Example 18.33 Revisited II
If x < 0, µ1 = 0, which is forbidden by the condition that at least one µi
must be non-zero.
Finally, if x = 0, y = 0, and any value of µ1 will satisfy the first order
equations. This yields the correct solution (x, y) = (0, 0).

x3 + y2 = 0

b
x∗
x
x=0

Figure 19.4.2: The solution is at the cusp, x∗ = (0, 0). There Df(x∗ ) = (0, 0),
forcing the multiplier µ0 = 0 because Df = (1, 0) 6= (0, 0).
 19. CONSTRAINED OPTIMIZATION II 39
19.4.4 Fritz John Theorem
This version of Theorem 19.4.1, including a multiplier on the objective,
allows for inequality constraints.7
Theorem 19.4.3. Let U ⊂ Rm . Suppose f : U → R and g : U → Rk
are C1 functions. Suppose further that x∗ is a local max of f under the
constraints
g(x) ≤ b.

Form the Lagrangian

k
X

L(x, λ0, . . . , λk ) = λ0f(x) − λj gj (x) − bj
j=1

with a multiplier λ0 for the objective function. Then there exist λ∗0, . . . , λ∗k
such that:
1. Dx L(x∗, λ∗0, . . . , λ∗k ) = 0.
2. λ∗1[g1 (x∗) − b1 ], . . . , λ∗k [gk (x∗) − bk ] = 0.
3. λ∗1 ≥ 0, . . . , λ∗k ≥ 0.
4. g(x∗) ≤ b.
5. λ∗0 = 0 or λ∗0 = 1, and
6. (λ∗0, . . . , λ∗k ) 6= 0.

7
The work of Karush, Kuhn, Tucker, and John founded the subject of nonlinear
programming, the study of nonlinear constrained optimization. Fritz John’s original
paper was Fritz John (1948), “Extremum problems with inequalities as subsidiary condi-
tions” in K.O. Friedrichs et al. (eds.), Studies and Essays, Courant Anniversary Volume,
Wiley/Interscience (Reprinted in: J. Moser (ed.): Fritz John Collected Papers vol. 2,
Birkhäuser, 1985, pp. 543-560).
40 MATH METHODS
19.4.5 More Constraint Qualification: Maximia
There are other constraint qualification conditions that can be used. The
following theorem collects some of the them without proof.
Theorem 19.4.4. Let U ⊂ Rm and suppose f : U → R and g : U → Rk
are C1 functions. Suppose also that x∗ is a local maximum of f under the
constraints
g(x) ≤ b.

Let ĝ(x) ∈ Rh denote the vector of binding constraints at x∗ , and suppose

ĝ obeys any of the following:
(a) NDCQ: The derivative Dĝ(x∗) has rank h.
(b) Karush-Kuhn-Tucker CQ: For any v ∈ Rm obeying Dĝ(x∗) v ≤ 0
there is an ε > 0 and a C1 curve α : [0, ε) → Rm such that:
(i) α(0) = x∗ ,
(ii) α′ (0) = v, and


(iii) gi α(t) ≤ bi for i = 1, . . . , k and all t ∈ [0, ε).
(c) Slater CQ: There is a ball V ⊂ Rm containing x∗ with the gi convex
on V and there exists a point z ∈ V with g(z) ≪ b.
(d) Concave CQ: The gi are concave functions.
(e) Linear CQ: The gi are linear functions.
Then we can set λ∗0 = 1 in Theorem 19.4.3.
The first constraint qualification condition is our old friend, the NDCQ.
The second was used by Kuhn and Tucker in their original work on
optimization with inequality constraints. The Slater condition applies
with convex constraints and requires that the constraint set have a non-
empty interior. The concave CQ ensures the constraint set is convex.
The last condition derives from linear programming models. It can be
considered a special case of the Concave CQ.
 19. CONSTRAINED OPTIMIZATION II 41
19.4.6 More Constraint Qualification: Minima
Some of the constraint qualification conditions work a bit differently for
minima. The way the constraints are written has been reversed in the
following theorem. Further, the KKTCQ, Slater CQ, and Concave (now
Convex) CQ are all modified.
Theorem 19.4.5. Let U ⊂ Rm and suppose f : U → R and g : U → Rk
are C1 functions. Suppose also that x∗ is a local minimum of f under the
constraints
g(x) ≥ b.

Let ĝ(x) ∈ Rh denote the vector of binding constraints at x∗ , and suppose

ĝ obeys any of the following:
(a) NDCQ: The derivative Dĝ(x∗) has rank h.
(b) Karush-Kuhn-Tucker CQ: For any v ∈ Rm obeying Dĝ(x∗) v ≥ 0
there is an ε > 0 and a C1 curve α : [0, ε) → Rm such that:
(i) α(0) = x∗ ,
(ii) α′ (0) = v, and


(iii) gi α(t) ≥ bi for i = 1, . . . , k and all t ∈ [0, ε).
(c) Slater CQ: There is a ball V ⊂ Rm containing x∗ with each gi concave
on V and there exists a point z ∈ V with g(z) ≫ b.
(d) Convex CQ: The gi are convex functions.
(e) Linear CQ: The gi are linear functions.
Then we can set λ∗0 = 1 in Theorem 19.4.3.

December 6, 2022

Copyright c 2022 by John H. Boyd III: Department of Economics, Florida In-

ternational University, Miami, FL 33199