一、4 支出最小化

4.1 支出最小化问题的求解∗ （MWG: 3.E; JR: 1.4.2; V: 7, 8, 9）

4.2 希克斯需求函数（对应）的性质 （MWG: 3.E; JR: 1.4.2, 1.4.3; V: 7, 8, 9）
4.3 支出函数的定义与性质 （MWG: 3.E; JR: 1.4.2, 1.4.3; V: 7, 8, 9）
4.4 支出函数与希克斯需求的关系 （将在一、5中讲）

3.E The Expenditure Minimization Problem

1. EMP
min p · x s.t. u(x) ≥ u
x≥0

▶ graphical illustration in Fig. 1

▶ Lagrangian function
X
L
L= pℓ xℓ + λ[u − u(x1 , · · · , xL )]
ℓ=1

F.O.C.:
∂u(x∗1 , · · · , x∗L )
pℓ − λ ≥ 0, with equality if x∗ℓ > 0, ℓ = 1, · · · , L
∂xℓ
u(x∗1 , · · · , x∗L ) ≥ u, with equality if λ∗ > 0

• Proposition 3.E.3.(ii) shows that, under its assumptions, u(x1 , · · · , xL ) ≥ u always hold with
equality, so we can simply replace the line above by u(x∗1 , · · · , x∗L ) = u.

▶ Cobb-Douglas example (assume that the target utility level u is strictly positive)

L = p1 x1 + p2 x2 + λ[u − xα 1−α
1 x2 ]

F.O.C.: (Recall that, with Cobb-Douglas utility function, the indifference curves corresponding
to strictly positive utility levels lie strictly within the first quadrant (i.e., they do not touch the
axes), so all x∗ℓ will be strictly positive.)

p1 − λα(x∗1 )α−1 (x∗2 )1−α = 0

p2 − λ(1 − α)(x∗1 )α (x∗2 )−α = 0

u − (x∗1 )α (x∗2 )1−α = 0

The solutions are usually denoted hℓ (p, u) and called the Hicksian, or compensated, demand
correspondence/function. In this example, we have
 1−α
αp2
h1 (p, u) = u
(1 − α)p1
 α
(1 − α)p1
h2 (p, u) = u
αp2

1
 Substituting these into the objective function, we get the value function, the expenditure function,

e(p, u) = p · h(p, u) = p1 h1 (p, u) + p2 h2 (p, u) = (α−α (1 − α)α−1 )pα 1−α

1 p2 u.

• Note that these formulas still give the correct values of h(p, u) and e(p, u) even when the target
utility level u is zero.

2. Proposition 3.E.1 on the relation between UMP and EMP

Suppose u(·) is a continuous utility function representing a locally non-satiated preference relation
≿ on the consumption set X = IRL
+ and that all prices are strictly positive p ≫ 0. Then,

(1) If x∗ is optimal in the UMP when wealth is w > 0, then x∗ is optimal in the EMP when the
required utility level is u(x∗ ). Moreover, the minimized expenditure level in the EMP is exactly
w.

(2) If x∗ is optimal in the EMP when the required utility level is u > u(0), then x∗ is optimal in
the UMP when wealth is p · x∗ . Moreover, the maximized utility level in the UMP is exactly u.

[That is, if a UMP and an EMP are dual problems, they have the same optimizers, and the
optimized objective value in one problem is equal to a key parameter in the other problem.]

• Proof of (1) by contradiction:

⋄ Suppose x∗ is not optimal in the EMP with required utility level u(x∗ ). Then, there exists an
x′ that brings at least a level of utility u(x∗ ) but costs strictly less than p · x∗ . In symbols, ∃ x′
such that p · x′ < p · x∗ and u(x′ ) ≥ u(x∗ ). Using p · x∗ ≤ w, we have p · x′ < p · x∗ ≤ w.

⋄ By local non-satiation, there exists some bundle x′′ very close to x′ that brings a strictly
higher level of utility than x′ but still costs less than w. That is, ∃ x′′ such that p · x′′ < w and
u(x′′ ) > u(x′ ).

⋄ But this implies that x′′ ∈ Bp,w and u(x′′ ) > u(x∗ ), contradicting the optimality of x∗ in the
UMP.

⋄ Thus, x∗ must be optimal in the EMP when the required utility level is u(x∗ ), and the minimized
expenditure level is p · x∗ . Finally, since x∗ solves the UMP when wealth is w, Walras’ Law implies
that p · x∗ = w.

⋄ The EMP is the ‘dual’ problem to the UMP.

• Proof of (2) by contradiction:

⋄ (Since u > u(0), we have x∗ ̸= 0, so p · x∗ > 0.)

2
 ⋄ Suppose that x∗ is not optimal in the UMP when wealth is p · x∗ . Then, there exists a bundle
x′ that costs no more than p · x∗ but brings a utility level strictly greater than u(x∗ ). In symbols,
∃ x′ such that p · x′ ≤ p · x∗ and u(x′ ) > u(x∗ ).

⋄ Now, scale down the bundle x′ a little bit to get x′′ = αx′ , where α ∈ (0, 1).

⋄ By continuity of u(·), if α is close enough to 1 so that x′′ is close enough to x′ , then x′′ will cost
strictly less than p · x∗ but still bring a utility level greater than u(x∗ ). In symbols, we can have
u(x′′ ) > u(x∗ ) and p · x′′ < p · x∗ .

⋄ This contradicts the optimality of x∗ in the EMP. So x∗ must be optimal in the UMP when
wealth is p · x∗ and the maximized utility level is u(x∗ ).

⋄ (A similar argument in Proposition 3.E.3(ii) will show that, if x∗ solves the EMP with required
utility level u, then u(x∗ ) = u.)

2.1 Implications/Restatements of the Proposition (refer to Fig. 1):

▶ Implication 1: Eq. 3.E.1

e(p, v(p, w)) = w and v(p, e(p, u)) = u

Loosely speaking, these mean the following: Given prices p, the minimum expenditure needed
to achieve the utility level v(p, w) is w, and the maximum utility that can be achieved with the
wealth level e(p, u) is u.

• These conditions imply that, for a fixed price vector p, e(p, ·) and v(p, ·) are inverses of one
another.

• This provides a method of deriving one function directly from the other—just invert the function.

• Example (2nd part of Exercise 3.E.8):

1−α
▷ the indirect utility function for the Cobb-Douglas utility function u(x1 , x2 ) = xα
1 x2 in
Example 3.D.2 is
v(p, w) = αα (1 − α)1−α p−α α−1
1 p2 w

▷ the expenditure function associated with this utility function is

e(p, u) = (α−α (1 − α)α−1 )pα 1−α

1 p2 u

• Quiz: Derive the expenditure function associated with the CES utility function from its indirect
utility function that we derived in section 3.D. Then, derive this expenditure function by solving
the EMP. Do you get the same answer?

3
▶ Implication 2: Eq. 3.E.4

h(p, u) = x(p, e(p, u)) and x(p, w) = h(p, v(p, w))

Informally, these mean the following: Given p and a level of wealth w = e(p, u), the utility
maximizing bundle x(p, w) is the same as the expenditure minimizing bundle h(p, u). Given p
and a target utility level u = v(p, w), the expenditure minimizing bundle h(p, u) is the same as
the utility maximizing bundle x(p, w).

• The first part, h(p, u) = x(p, e(p, u)), explains why h(p, u) is a compensated demand corre-
spondence: As prices vary, h(p, u) gives precisely the level of demand that would arise if the
consumer’s wealth were simultaneously adjusted to keep her utility level at u.

• Fig. 4. ∆wHicks = e(p′ , u) − w.

• This type of wealth compensation is known as Hicksian wealth compensation.

• As prices change, h(p, u) keeps utility constant by adjusting wealth, while x(p, w) keeps wealth
fixed but allows utility to vary.

3. Proposition 3.E.2 on the value function, the expenditure function e(p, u)

▶ expenditure function e(p, u) = min p · x s.t. u(x) ≥ u, where u is a continuous utility function
x
representing a locally non-satiated preference relation defined on X = IRL
+

▶ properties:

(1) homogeneous of degree one in p : e(αp, u) = αe(p, u).

   
Proof: e(αp, u) = min αp · x s.t. u(x) ≥ u = α min p · x s.t. u(x) ≥ u = αe(p, u)
x x

intuition: ∵ The constraint set does not change with α, nor do the contours of the objective
function ∴ minimizing αp · x on the unaltered constraint set will lead to the same optimal solution
x∗ as minimizing p · x.

(2) strictly increasing in u and non-decreasing in p

• if u′ > u, then e(p, u′ ) > e(p, u)

Proof by contradiction: Let x and x′ denote optimal choices for required utility levels u and u′ .
If u′ > u and p · x ≥ p · x′ > 0, then, by the continuity of u(·), there exist x
e = αx′ and α ∈ (0, 1)
x) ≥ u and p · x
close to 1 such that u(e e < p · x. This contradicts x being optimal for the EMP with
required utility u.

• if p′ℓ ≥ pℓ and p′k = pk for all k ̸= ℓ, then e(p′ , u) ≥ e(p, u)

4
 first note that the price vectors p′ = (p1 , · · · , p′ℓ , · · · , pL ) and p = (p1 , · · · , pℓ , · · · , pL ) differ only
in their ℓth element

Proof: e(p′ , u) = p′ · x′ ≥ p · x′ ≥ p · x = e(p, u).

(3) concave in p : e(αp + (1 − α)p′ , u) ≥ αe(p, u) + (1 − α)e(p′ , u), α ∈ [0, 1].

Proof:

Let p′′ ≡ αp + (1 − α)p′ . Suppose that x′′ is optimal for the EMP with prices p′′ and required
utility u.

Then, we have

e(p′′ , u) = p′′ · x′′ = αp · x′′ + (1 − α)p′ · x′′ ≥ αe(p, u) + (1 − α)e(p′ , u),

where the last inequality follows because u(x′′ ) ≥ u and therefore, by the definition of the expen-
diture function, p · x′′ ≥ e(p, u) and p′ · x′′ ≥ e(p′ , u).

An aside:
min [f (x) + g(x)] ≥ min f (x) + min g(x)
x x x

max [f (x) + g(x)] ≤ max f (x) + max g(x)

x x x

so an alternative way of writing the expressions in the proof above is:

min [αp + (1 − α)p′ ] · x s.t. u(x) ≥ u

x

= min αp · x + (1 − α)p′ · x s.t. u(x) ≥ u

x

≥ min αp · x s.t. u(x) ≥ u + min (1 − α)p′ · x s.t. u(x) ≥ u

x x

intuition for the concavity result (see Figure 3.E.2):

• When prices change but the consumer cannot adjust consumption x, expenditure p · x is a linear
function of p.

• If prices change and the consumer can adjust consumption x, the resulting expenditure will be
no greater than the level specified by the linear function p · x above.

3.1 Example: Verify the concavity in prices of the expenditure function associated with the Cobb-
Douglas utility function.

▶ We need two theorems from M.C and M.D of MWG.

• The first part of Theorem M.C.2 tells us that f : A → IR is concave if and only if the Hessian
matrix H ≡ D2 f (x) is negative semidefinite for every x ∈ A.

5
 • Part two of Theorem M.D.2 provides a method for checking whether H is negative semidefinite.
Since H will be a 2 × 2 symmetric matrix, it satisfies the requirement of symmetry in part
two of this theorem. Then, according to the theorem, H is negative semidefinite if and only if
(−1)r |r Hrπ | ≥ 0 for every r = 1, 2 and for every permutation π of the indices {1, 2}, that is, for
π1 = (1, 2) and π2 = (2, 1).

• Theorem 16.2 of Simon and Blume (1994) Mathematics for Economists provides a simpler,
alternative statement of MWG Theorem M.D.2(ii) above. For negative semidefiniteness of a
symmetric matrix, we just need all its principal minors of odd order to be non-positive (≤ 0)
and all its principal minors of even order to be non-negative (≥ 0). See the images of the relevant
pages.

▶ The calculations below makes use of MWG Theorem M.D.2(ii). Recall that, for the Cobb-Douglas
1−α
utility function u(x1 , x2 ) = xα
1 x2 , the associated expenditure function is given by

e(p, u) = (α−α (1 − α)α−1 )pα 1−α

1 p2 u.

To focus on the concavity in (p1 , p2 ), we can ignore the positive constant in this exercise, (α−α (1−
α)α−1 )u, and focus on the concavity in p of the following function:

1−α
f (p) = pα
1 p2 .

▶ First, calculate the first- and second-order derivatives of f (p1 , p2 ) :

−α
f1 = αpα−1
1 p1−α
2 , f2 = (1 − α)pα
1 p2

f11 = α(α − 1)pα−2

1 p1−α
2 , f12 = f21 = α(1 − α)pα−1
1 p−α
2 , f22 = −α(1 − α)pα −α−1
1 p2

▶ So the Hessian matrix is  

f f12
H = 11
f21 f22

▶ H has two permutations, π1 = (1, 2) and π2 = (2, 1) :

 
f f21
H π1
= H itself, H π2
= 22
f12 f11

▶ For permutation π1 , we have that, for r = 1,

|1 H1π1 | = f11

and, for r = 2,
|2 H2π1 | = |H| = f11 f22 − f21 f12 .

6
 We need to check whether

(−1)1 |1 H1π1 | = −f11 ≥ 0 and (−1)2 |2 H2π1 | = f11 f22 − f21 f12 = f11 f22 − f12
2
≥ 0.

▶ For permutation π2 , we have that, for r = 1,

|1 H1π2 | = f22

and, for r = 2,
|2 H2π2 | = |H| = f11 f22 − f12 f21 .

We need to check whether

(−1)1 |1 H1π2 | = −f22 ≥ 0 and (−1)2 |2 H2π2 | = f11 f22 − f12 f21 = f11 f22 − f12
2
≥ 0.

▶ In our example, we have f11 < 0 and f22 < 0, because α ∈ (0, 1). Moreover,
−α−1
f11 f22 − f12
2
= [α(α − 1)pα−2
1 p1−α
2 ][−α(1 − α)pα
1 p2 ] − [α(1 − α)p1α−1 p−α
2 ]
2

= α2 (α − 1)2 [p2α−2
1 p−2α
2 − p2α−2
1 p−2α
2 ] = 0.

▶ Thus, the criterion in part (ii) of Theorem M.D.2 is indeed satisfied, and f (p1 , p2 ) is concave in
(p1 , p2 ). Then so is e(p, u).

▶ Now, let us use the method in the Simon and Blume (1994) book. The method is for checking
negative semidefiniteness of a symmetric matrix. So we need to first make sure that the matrix
is symmetric. This holds in our case because of Young’s theorem: f12 = f21 .

▶ There are two principal submatrices of odd order (of order 1): (f11 ) and (f22 ). So the two principal
minors of odd order are simply f11 and f22 . We see from the calculations above that they are
both negative.

▶ There is only one principal submatrix of even order, that of order 2:

 
f11 f12
.
f21 f22

So the only principal minor of even order is f11 f22 − f12

2
, which is zero and therefore non-negative
as shown above.

▶ Therefore, according to Theorem 16.2 of Simon and Blume (1994), the Hessian matrix is negative
semidefinite, and the expenditure function is concave in p.

4. Properties of the optimal choice, the Hicksian demand h(p, u)

7
 ((1) through (3) are Proposition 3.E.3. (4) is Proposition 3.E.4. Both propositions assume that
u is a continuous utility function representing a locally non-satiated preference relation defined
on X = IRL
+ . Proposition 3.E.4 further assumes that h(p, u) is a function.)

(1) homogeneous of degree zero in p : h(αp, u) = h(p, u)

[similar idea as proof of e(p, u) being HD1 in p in Proposition 3.E.2] The optimal vector that
minimizes p · x subject to u(x) ≥ u is the same as the optimal vector that minimizes αp · x subject
to the same constraint, for any α > 0.

[can also show this by combining Prop. 3.E.2.(i), Prop. 3.G.1 and Theorem M.B.1 in math
appendix M.B]

(2) no excess utility: for any x ∈ h(p, u), u(x) = u

Proof by contradiction: Suppose x ∈ h(p, u) and u(x) > u. By the continuity of u(·), there exists
y = αx with α close to 1 such that u(y) ≥ u and p · y < p · x. This contradicts x being optimal
for the EMP with parameters p and u.

(3) convexity/uniqueness: If ≿ is convex, h(p, u) is a convex set. If ≿ is strictly convex, h(p, u)

consists of a single element.

Proof: Suppose x ∈ h(p, u), x′ ∈ h(p, u), x ̸= x′ , and p · h(p, u) = e. Denote x′′ ≡ αx + (1 − α)x′ .

Convexity: Since ≿ is convex (so u(x) is quasiconcave) and u(x) = u(x′ ) = u, we have u(x′′ ) ≥ u.
Moreover, since p · x = p · x′ = p · h(p, u) = e, we have p · x′′ = p · (αx + (1 − α)x′ ) = e. Therefore,
x′′ ∈ h(p, u).

Uniqueness: Suppose x ̸= x′ . Since ≿ is strictly convex (so u(x) is strictly quasiconcave) and
u(x) = u(x′ ) = u, we have u(x′′ ) > u. Besides, we have shown that p · x′′ = e. By the continuity
of u(·), there exists y = αx′′ with α close to 1 such that u(y) ≥ u and p · y < e. This contradicts
the assumption that x and x′ are elements of h(p, u). So x ̸= x′ cannot be true.

(4) Hicksian demand function satisfies the compensated law of demand: For all p′ and p′′ ,

(p′′ − p′ ) · [h(p′′ , u) − h(p′ , u)] ≤ 0.

That is, demand and price move in opposite directions for price changes that are accompanied
by Hicksian wealth compensation.

Proof for the case of single-valued Hichsian demand: By the definition of h(p, u), we have

p′′ · h(p′′ , u) ≤ p′′ · h(p′ , u),

p′ · h(p′′ , u) ≥ p′ · h(p′ , u).

8
 Then, subtracting these inequalities gives the result.

(A special case: Suppose that only pℓ changes. Then, (p′′ℓ − p′ℓ )[hℓ (p′′ , u) − hℓ (p′ , u)] ≤ 0. That is,
for compensated price changes, own price effects are non-positive.)

In contrast, Warasian demand x(p, w) need not satisfy the law of demand (e.g., Giffen goods).

4.1 Example with Cobb-Douglas utility function: Verify the first three properties of h(p, u) listed
above
1−α
We found above that, when u(x1 , x2 ) = xα
1 x2 , we have
 1−α
αp2
h1 (p, u) = u
(1 − α)p1
 α
(1 − α)p1
h2 (p, u) = u
αp2
e(p, u) = (α−α (1 − α)α−1 )pα 1−α
1 p2 u

(1) Let t be some positive constant. We have

 1−α  1−α
αtp2 αp2
h1 (tp, u) = u= u = h1 (p, u)
(1 − α)tp1 (1 − α)p1
 α  α
(1 − α)tp1 (1 − α)p1
h2 (tp, u) = u= u = h2 (p, u)
αtp2 αp2

So h1 (p, u) and h2 (p, u) are HD0 in p.

1−α
(2) Substituting h1 (p, u) and h2 (p, u) in the utility function xα
1 x2 , we get
" 1−α #α  α 1−α
αp2 (1 − α)p1
u(x1 , x2 ) = u u ,
(1 − α)p1 αp2

which is exactly equal to u. So the no excess utility property is verified.

(3) We know that the Cobb-Douglas preference is strictly convex when both x1 and x2 are strictly
positive. According to property (iii) above, h(p, u) must be a singleton. From the expressions of
h1 (p, u) and h2 (p, u), we see that this is indeed the case.

5. One more example

(1) Derive the Hicksian demand functions, the expenditure function and the indirect utility function
√ √
for the utility function u(x1 , x2 ) = x1 + 2 x2 . (Assume that p1 > 0, p2 > 0 and the target
utility level u > u(0, 0).)

Suggested answer: The expenditure-minimization problem is

√ √
min p 1 x 1 + p2 x 2 , s.t. x1 + 2 x2 ≥ u
{x1 ,x2 }

9
The Lagrangian is
√ √
L = p1 x1 + p2 x2 + λ(u − x1 − 2 x2 )

The first-order conditions are

∂L 1 −1 ∂L
= p1 − λx1 2 ≥ 0, x1 ≥ 0, x1 =0
∂x1 2 ∂x1

∂L −1 ∂L
= p2 − λx2 2 ≥ 0, x2 ≥ 0, x2 =0
∂x2 ∂x2
∂L √ √ ∂L
= u − x1 − 2 x2 ≤ 0, =0 λ ≥ 0, λ
∂λ ∂λ
√ √
Since the MRS for this utility function, −M U1 /M U2 = − x2 /(2 x1 ), is zero when x2 = 0 and
infinity when x1 = 0, p1 and p2 are strictly positive (so −∞ < −p1 /p2 < 0), and the target utility
satisfies u > u(0, 0), we know that an optimum must be interior, that is x1 > 0 and x2 > 0 in
an optimum. Since u(x1 , x2 ) is strictly increasing (so preference is strongly monotone and hence
locally non-satiated), we know that the target utility constraint must be binding in an optimum.
Therefore, despite the a priori mathematical possibilities of boundary solutions and non-binding
constraint, economic reasoning helps us rule them out, and all of the first-order conditions above
will in fact be equalities:
∂L 1 −1
= p1 − λx1 2 = 0 (1)
∂x1 2
∂L −1
= p2 − λx2 2 = 0 (2)
∂x2
∂L √ √
= u − x1 − 2 x2 = 0 (3)
∂λ

Given this, we can easily solve for the Hicksian demand functions as follows. First, from (1) and
(2), we can get
 2
2p1
x2 = x1
p2
Substitution of this into (3) yields
s 2
√ 2p1
x1 + 2 x1 − u = 0
p2

Thus  2  2
p2 u 2p1 u
h1 (p1 , p2 , u) = , h2 (p1 , p2 , u) = .
4p1 + p2 4p1 + p2
Substituting these solutions into the objective function, we get the expenditure function
 2  2
p2 u 2p1 u p 1 p2
e(p1 , p2 , u) = p1 + p2 = u2 .
4p1 + p2 4p1 + p2 4p1 + p2

10
 By Proposition 3.E.1 and Equations 3.E.1, the indirect utility function can be found by inverting
the expenditure function. Writing

e(p1 , p2 , v(p1 , p2 , w)) = w,

that is,
p1 p2
[v(p1 , p2 , w)]2 = w,
4p1 + p2
and inverting, we have s
(4p1 + p2 )w
v(p1 , p2 , w) = .
p1 p2

(2) We can see from part (1) that u(x) is homogeneous of degree 1/2 in x and e(p, u) is homogeneous
of degree 2 in u. Is it always true that, for any u(x) that is homogeneous of degree 1/2 in x, the
corresponding expenditure function e(p, u) will be homogeneous of degree 2 in u? Why or why
not?

Suggested answer: Yes, the statement is true. This can be shown as follows. Let t be a positive
constant. Then,
   
e(p, tu) = min p · x s.t. u(x) ≥ tu = min p · x s.t. t−1 u(x) ≥ u
x x

   
−2 −2 −2
= min p · x s.t. u(t x) ≥ u = min t (p · t
2
x) s.t. u(t x) ≥ u
x x
 
=t 2
min (p · x
e) s.t. u(e
x) ≥ u = t2 e(p, u),
e
x

e ≡ t−2 x.
where x

Recommended Exercises: 3.E: 4, 5, 6, 7, 8, 9, 10

11