Lecture 4
0.1 Recovering the utility function from an expenditure
function
1. We saw that any expenditure function must satisfy the properties 1 to 6
of Thm 1.7. We now show that the converse is also true: any function
E : <n++ <+ ! <+ that satis…es properties 1 through 6 in Thm 1.7
and is di¤erentiable can be interpreted as the expenditure function of an
increasing, quasiconcave utility function.
2. This is known as the duality problem— the utility function and the expen-
diture functions are “duals,”meaning that despite the apparent di¤erence
they contain the same information about the consumer so that if we know
one, we know the other. We have used the better-than set to derive the ex-
penditure function. It is also possible to do it the other way round— using
the the expenditure function to derive the utility function.
3. Given e (p; u) (de…ned for all p >> 0, and all u), we call the hyperplane
px = e (p; u)
a supporting hyperplane of u.
4. For a …xed u, each hyperplane is associated with a price vector p. Take
intersection over all p >> 0, we have
A (u) x 2 <n+ jp:x e (p; u) for all p >> 0 : (1)
5. Theorem 3.H.1 (Mas-Colell). Let U be a connected subset of <. Suppose
e : <n++ <+ ! <+ is continuous, strictly increasing, and unbounded in
u, di¤erentiable, strictly increasing, concave and homogeneous of degree
one in p (recall that these are all properties of expenditure functions of
a quasi-concave, strictly increasing utility function). Then the function
u : <n+ <+ given by
u (x) max fujx 2 A (u)g (2)
(A (u) as de…ned in (1)) is increasing and quasiconcave, and e is the ex-
penditure function of u; that is,
e (p; u) = min p:x s.t. u (x) u:
x
6. Proof:
(a) Step 0: A (u) is non-empty. For any p; p
e (p; u) e (p ; u) + re (p ; u) (p p )
= p re (p ; u) + re (p ; u) (p p )
= re (p ; u) p:
1
The inequality follows from the concavity of e, and the equality fol-
lows from the fact that e is homogeneous of degree one in p. Finally,
re (p ; u) 0 as e increases in p. It follows re (p ; u) 2 A (u).
(b) Step 1: u is well-de…ned (meaning that the max exists). Since e (p; u)
is strictly increasing and unbounded in u, the set fu 0jx 2 A (u)g
is bounded from above. Let u denote sup fu 0jx 2 A (u)g, and
u1 ; u2 ; :::be a sequence that converges to u , with ui 2 fu 0jx 2 A (u)g
for all i. Suppose u 2 = fu 0jx 2 A (u)g. Then there exists some
p >> 0 such that p x < e (p ; u ). But since e is continuous in u,
p x < e (p ; un ) for n su¢ ciently large, a contradiction.
(c) Step 2: u is increasing. Consider x1 x2 . Since px1 px2 for all p,
x2 2 A (u) whenever x1 2 A (u). Hence, u x2 must be as least as
great as u x1 .
(d) Step 3: u is quasiconcave. We want to show that for any x1 and x2
and any 2 [0; 1] ;
u x1 + (1 ) x2 min u x1 ; u x2 :
In order words, we want to show that the better than set of x is
convex. Without loss of generality, suppose u x1 u x2 . Since e
is increasing in u, we have
e p; u x1 + (1 ) e p; u x2 e p; u x1 :
By de…nition of u
px1 e p; u x1
px2 e p; u x2
It then follows that
p x1 + (1 ) x2 e p; u x1 + (1 ) e p; u x2
e p; u x1 :
which implies that u x1 + (1 ) x2 u x1 .)
(e) We now proceed to show that
e (p; u) = min p:x s.t. u (x) u:
x
Note that for all x such that u (x) u;
p:x e (p; u (x)) (by defn of u (x) ),
e (p; u) (e is increasing in u).
Hence,
e (p; u) min p:x s.t. u (x) u:
x
2
To complete the proof we will show that
e (p; u) min p:x s.t. u (x) u:
x
Since e is concave in p, for any p, p0 and u
e (p; u) + rp e (p; u) (p0 p) e (p0 ; u) :
Since e is h.o.d 1 in p,
e (p; u) = p:rp e (p; u)
by the Euler equation. Combining the two equations, we have for all
p0
p0 rp e (p; u) e (p0 ; u) :
Since e is increasing in p, rp e (p; u) 2 <n+ . It follows that, rp e (p; u) 2
A (u) and, hence, u (rp e (p; u)) u. This implies that
e (p; u) = p:rp e (p; u) min p:x s.t. u (x) u:
x
7. If e is the expenditure function of a strictly convex preference relation,
then the procedure above will recover the exact utility function. IN this
case every x0 not belonging to the better-than set is excluded from A (u).
By the separating hyperplane theorem, there exists a hyperplane that
separates x0 from the better than set. If a hyperplane separates x0 from
the better-than set, so does the supporting hyperplane with the same
slope.
8. If e is the expenditure function of a non-convex preference relation. (In this
case e will not be di¤erentiable.) The above proof will still work with minor
changes. But the A (u) recovered will be the convex hull of the underlying
better-than set. In this sense, convexity can be assumed without loss of
generality— any rational consumer choice is consistent with some convex
preference relation. Intuitively, consumption bundles on the non-convex
part of an indi¤erent curve are never chosen, and hence information about
them are not recorded in the expenditure function.
Finding u Algebraically
1=r
1. Suppose e = u (pr1 + pre ) . How can we …nd u? The above discussion
shows that we can construct u by solving the following problem:
u (x) = max u s.t. px e (p; u) 8p >> 0: (3)
u
(Note that x is …xed in this problem.) This problem is hard to solve at
the …rst glance because it involves an in…nite number of constraints. But
we can transform this problem into one that is solvable.
3
2. De…ne
v (p; y) = max u s.t. e (p; u) y:
u
Solving this problem is easy as e is increasing in u. Just set e (p; u) = y.
1=r
For example, if e = u (pr1 + pre ) , then
1=r
v = y (pr1 + pre ) :
3. Instead of solving (3), we …rst …x p >> 0 and consider the problem:
max u s.t. px e (p; u) : (4)
u
4. By construction, the solution to this problem is v (p; px). Since we consider
in (4) only a subset of constraints in (3) (as a result, any feasible solution
to (3) is feasible in (4)) for any p >> 0; v (p; px) u (x). Hence
min v (p; px) u (x) :
p>>0
5. On the other hand, since e is increasing in u; for all p >> 0
px e (p; v (p; px)) e p; min v (p; px) :
p
Hence, minp v (p; px) is a feasible solution to (3); therefore,
u (x) min v (p; px) :
p
6. It follows that
u (x) = min v (p; px) : (5)
p
This problem can be solved readily.
0.1.1 An alternative interpretation
1. We do not explicitly cover production theory in this class because it is
more or less the same as consumer theory. Whereas a consumer con-
sumes goods to produce utility, a …rm consumes inputs to produce out-
puts. Utility maximization is equivalent to maximizing pro…t given a …xed
input budget, and cost minimization is equivalent to minimizing produc-
tion cost given an output target. So what we just did is equivalent to
recovering the production function from the cost function of a …rm. This
is an important technique because cost data are more readily available
than production data.
4
0.2 A Brief Note on Integrability
1. We have shown how to construct a utility function from an expenditure
function. In fact we carry out the analysis further— starting with a de-
mand function x (p; y) we can cover the expenditure function.
2. Speci…cally, given x (p; y) we want to know whether there exists e (p; u)
such that for i = 1; :::; n;
@e (p; u)
= xi (p; e (p; u)) : (6)
@pi
This is called the integrability problem because we observe the derivatives
and try to integrate them back into e.
3. In e¤ect we are trying to solve a system of partial di¤erential equations. It
turns out, quite surprisingly, that this could be done if and only if x (p; y)
satis…es budget balancedness, and its substitution matrix is symmetry and
negative semide…nite.
4. This, combined with the fact that we can recover the utility function from
e, means that any demand function x that satis…es budget balancedness
and possesses a symmetric and negative semide…nite substitution matrix is
rationalizable by a strictly increasing quasiconcave utility function. Since
we have already shown that x generated by a strictly increasing quasi-
concave utility function must possess these properties. Thus, budget bal-
ancedness, symmetry and negative semide…niteness are the only conclu-
sions we can draw if all we know is that the consumer has a quasiconcave,
strictly increasing utility function.
5. This is an important result because it allows economists to directly write
down demand functions that possess additional properties that are de-
sirable for empirical estimation. As long as the function satis…es budget
balancedness, symmetry and negative semide…niteness, we know that it is
consistent with rational behavior. There is thus no need to start explicitly
with a utility function. The problem with the latter approach is that it
is di¢ cult to make sure that the demand function generated by a utility
function possesses the properties we want.
