REVEALED PREFERENCE 131

Adding together and using the budget constraint yields

Substitute back into the first-order conditions to find

These are the choices of (pl,p2) that minimize indirect utility. Now
substitute these choices into the indirect utility function:

a b
u ( x ~ , x=
~ )-aln - bln
(a + b)xl (a + b12.
+ +
= a In XI b In x2 constant.

This is the familiar Cobb-Douglas utility function.

8.7 Revealed preference

In our study of consumer behavior we have taken preferences as the prim-

itive concept and derived the restrictions that the utility maximization
model imposes on the observed demand functions. These restrictions are
basically the Slutsky restrictions that the matrix of substitution terms be
symmetric and negative semidefinite.
These restrictions are in principle observable, but in practice they leave
something to be desired. After all, who has really seen a demand function?
The best that we may hope for in practice is a list of the choices made un-
der different circumstances. For example, we may have some observations
on consumer behavior that take the form of a list of prices, p t , and the
associated chosen consumption bundles, xt for t = 1,.. . , T. How can we
tell whether these data could have been generated by a utility-maximizing
consumer?
We will say that a utility function rationalizes the observed behavior
> >
(p t , x t ) for t = 1,.. - , T if u(x t ) u(x) for all x such that ptxt p t x.
That is, u(x) rationalizes the observed behavior if it achieves its maximum
value on the budget set at the chosen bundles. Suppose that the data were
generated by such a maximization process. What observable restrictions
must the observed choices satisfy?
Without any assumptions about u(x) there is a trivial answer to this
question, namely, no restrictions. For suppose that u(x) were a constant
function, so that the consumer was indifferent to all observed consumption
132 CHOICE (Ch. 8)

bundles. Then there would be no restrictions imposed on the patterns of

observed choices: anything is possible.
To make the problem interesting, we have to rule out this trivial case.
The easiest way to do this is to require the underlying utility function to be
locally nonsatiated. Our question now becomes: what are the observable
restrictions imposed by the maximization of a locally nonsatiated utility
function?
First, we note that if ptxt > p t x, then it must be the case that u ( x t ) 2
u ( x ) . Since xt was chosen when x could have been chosen, the utility of xt
must be at least as large as the utility of x . In this case we will say that xt is
directly revealed preferred to x , and write x t R D x. As a consequence of
this definition and the assumption that the data were generated by utility
maximization, we can conclude that "xtRDximplies u ( x t )2 u(x)."
Suppose that ptxt > p t x. Does it follow that u ( x t ) > u(x)? It is not
hard to show that local nonsatiation implies this conclusion. For we know
from the previous paragraph that u ( x t ) 2 u ( x ) ;if u ( x t ) = u ( x ) , then by
local nonsatiation there would exist some other x' close enough to x so that
ptxt > ptx' and u(x') > u ( x ) = u ( x t ) .This contradicts the hypothesis of
utility maximization.
If ptxt > p t x, we will say that xt is strictly directly revealed pre-
ferred t o x and write x t P D x .
Now suppose that we have a sequence of such revealed preference com-
parisons such that xt R ~ xxjRDxk1 ~ , . . . , x n RDx. In this case we will say
that xt is revealed preferred t o x and write x t Rx. The relation R is
sometimes called the transitive closure of the relation R ~If .we assume
that the data were generated by utility maximization, it follows that "xtRx
>
implies u ( x t ) u(x)."
Consider two observations xt and x". We now have a way to determine
whether u(xt) 2 u ( x S )and an observable condition t o determine whether
u ( x S )> u ( x t ) . Obviously, these two conditions should not both be satis-
fied. This condition can be stated as the

GENERALIZED AXIOM OF REVEALED PREFERENCE. If

xt is revealed preferred to x S , then x" cannot be strictly directly revealed
preferred to xt .

Using the symbols defined above, we can also write this axiom as

GARP. xt R x s implies not xS pD x t . I n other words, xt R x S implies

p5xS 5 psxt.

As the name implies, GARP is a generalization of various other revealed

preference tests. Here are two standard conditions.
 SUFFICIENT CONDITIONS FOR MAXIMIZATION 133

WEAK AXIOM OF REVEALED PREFERENCE (WARP). If

xt R D x h n d x t is not equal to x S , then it is not the case that x s R D x t .

STRONG AXIOM OF REVEALED PREFERENCE (SARP). If

x t R x s and x t is not equal to x S , then it is not the case that x s R x t .

Each of these axioms requires that there be a unique demand bundle at

each budget, while GARP allows for multiple demanded bundles. Thus,
GARP allows for flat spots in the indifference curves that generated the
observed choices.

8.8 Sufficient conditions for maximization

If the data (p t , x t ) were generated by a utility-maximizing consumer with

nonsatiated preferences, the data must satisfy GARP. Hence, GARP is an
observable consequence of utility maximization. But does it express all the
implications of that model? If some data satisfy this axiom, is it necessarily
true that it must come from utility maximization, or at least be thought
of in that way? Is GARP a suficient condition for utility maximization?
It turns out that it is. If a finite set of data is consistent with GARP,
then there exists a utility function that rationalizes the observed behavior-
i.e., there exists a utility function that could have generated that behavior.
Hence, GARP exhausts the list of restrictions imposed by the maximization
model.
The following theorem is the nicest way to state this result.

Afriat's theorem. Let (p t , x t ) for t = 1 , . . . ,T be a finite number of

observations of price vectors and consumption bundles. Then the following
conditions are equivalent.

(1) There exists a locally nonsatiated utilzty function that rationalizes the
data;

(2) The data satisfy GARP;

(3) There exzst positzve numbers (u t , At) for t = 1, . . . , T that satisfy the
Afmat inequalitzes:

us 5 ut + Atpt(xS- x t ) for all t , s ;

(4) There exists a locally nonsatiated, continuous, concave, monotonic util-

ity function that rationalizes the data.
134 CHOICE (Ch 8)

Proof. We have already seen that (1) implies ( 2 ) . The proof that ( 2 )
implies ( 3 ) is omitted; see Varian (1982a) for the argument. The proof
that (4) implies ( 1 ) is trivial. All that is left is the proof that (3) implies
(4).
We establish this implication consfructively by exhibiting a utility func-
tion that does the trick. Define

>
Note that this function is continuous. As long as pt 0 and no pt = 0, the
function will be locally nonsatiated and monotonic. It is also not difficult
to show that it is concave. Geometrically, this function is just the lower
envelope of a finite number of hyperplanes.
We need to show that this function rationalizes the data; that is, when
prices are pt, this utility function achieves its constrained maximum at x t .
F i t we show that u ( x t )= ut. If this were not the case, we would have

But this violates one of the Afriat inequalities. Hence, u ( x t )= ut.

>
Now suppose that p s x s p S x. It follows that

>
This shows that u ( x s ) u ( x ) for all x such that p S x 5 p s x S . In other
words, U ( X ) rationalizes the observed choices. I

The utility function defined in the proof of Afriat's theorem has a nat-
ural interpretation. Suppose that u ( x ) is a concave, differentiable utility
function that rationalizes the observed choices. The fact that u ( x ) is dif-
ferentiable implies it must satisfy the T first-order conditions

The fact that u ( x ) is concave implies that it must satisfy the concavity
conditions
+
u ( x t )5 u ( x S ) D U ( X ' ) ( X ~ - x S ) . (8.4)
Substituting from (8.3) into (8.4), we have

+
u ( x t ) 5 u ( x S ) XSpS(xt- x*),

Hence, the Afriat numbers ut and At can be interpreted as utility levels

and marginal utilities that are consistent with the observed choices.
The most remarkable implication of Afriat's theorem is that ( 1 ) implies
(4): if there is any locally nonsatiated utility functionat all that rationalizes
 COMPARATIVE STATICS USING REVEALED PREFERENCE 135

the data, there must exist a continuous, monotonic, and concave utility
function that rationalizes the data. This is similar t o the observation made
in Chapter 6, page 83, where we showed that if there were nonconvex
parts of the input requirement set, no cost minimizer would ever choose to
operate there.
The same is true for utility maximization. If the underlying utility func-
tion had the "wrong" curvature at some points, we would never observe
choices being made at such points because they wouldn't satisfy the right
second-order conditions. Hence market data do not allow us t o reject the
hypotheses of convexity and monotonicity of preferences.

8.9 Comparative statics using revealed preference

Since GARP is a necessary and sufficient condition for utility maximization,

it must imply conditions analogous to comparative statics results derived
earlier. These include the Slutsky decomposition of a price change into the
income and the substitution effects and the fact that the own substitution
effect is negative.
Let us begin with the latter result. When we consider finite changes
in a price rather than just infinitesimal changes, there are two possible
definitions of the compensated demand. The first definition is the natural
extension of our earlier definition-namely, the demand for the good in

of utility. That is, the value of the compensated demand for good i when
+ +
prices change from p to p A p is just z , ( p A p , m Am) + -
question if we change the level of income so as to restore the original level

x,(p+
A p , e ( p + A p , u)), where u is the original level of utility achieved at (p, m).
This notion of compensation is known as the Hicksian compensation.
The second notion of compensated demand when prices change from p
+
to p A p is known as the Slutsky compensation. It is the level of
demand that arises when income is changed so as t o make the original
level of consumptzon possible. This is easily described by the following
equations. We want the change in income, Am, necessary to allow for the
old level of consumption, x ( p , m), to be feasible at the new prices, p + A p .
That is
+ +
( p A p ) x ( p , m) = m Am.
Since p x ( p , m) = m, this reduces to A p x ( p , m) = Am.
The difference between the two notions of compensation is illustrated in
Figure 8.6. The Slutsky notion is directly measurable without knowledge of
the preferences, but Hicksian notion is more convenient for analytic work.
For infinitesimal changes in price there is no need to distinguish between
the two concepts since they coincide. We can prove this simply by exam-
ining the expenditure function. If the price of good j changes by dp,, we
need t o change expenditure by (de(p, u)/dp,)dp, t o keep utility constant.
 136 CHOICE (Ch. 8)

If we want to keep the old level of consumption feasible, we need to change

income by xjdpj. By the derivative property of the expenditure function,
these two magnitudes are the same.

Slutsky
compensation I
Htcks
compensation

GOOD 1

Figure Hicks a n d Slutsky compensation. Hicks compensation is

8.6 an amount of money that makes the original level of utility
affordable. Slutsky compensation is an amount of money that
makes the original consumption bundle achievable.

Whichever definition you prefer, we can still use revealed preference to

prove that "the compensated own-price effect is negative." Suppose we
consider the Hicksian definition. We start with a price vector p and let
+
x = x(p, m) be the demanded bundle. The price vector changes to p A p ,
+
and the compensated demand, therefore, changes to x ( p A p , m Am), +
where Am is the amount necessary to make x ( p + A p , m + Am) indifferent
to X(P, m).
+ +
Since x(p, m) and x ( p A p , m Am) are indifferent to each other,
neither can be strictly directly revealed preferred to the other. That is, we
' musthave
~x(~,m)<px(p+Ap,rn+Am)
(P + AP)X(P + AP, m + Am) I: ( p + Ap)x(p,m).
Adding these inequalities together, we have

Letting Ax = x ( p + A p , m + Am) - x(p, m), this becomes

 THE DISCRETE VERSION OF THE SLUTSKY EQUATION 137

Suppose that only one price has changed so that A p = (0,. . . , Api,.. . ,0).
Then this inequality implies that xi must change in the opposite direction.
We now turn t o the Slutsky definition. We keep the same notation as
before, but now interpret Am as the change in income necessary to make
the old consumption bundle affordable. Since x(p,m ) is thus by hypothesis
+
a feasible level of consumption a t p A p , the bundle actually chosen at
+
p A p cannot be revealed worse than x ( p ,m ). That is,
P X(P, + A p , m + Am).
m) I P x(p
Since ( p+ A p ) x ( p+ A p , m + Am) = ( p + A p ) x ( p ,m ) by construction
of Am, we can subtract this equality from the above inequality to find

just as before.

8.1 0 The discrete version of the Slutsky equation

We turn now to the task of deriving the Slutsky equation. We derived this
equation earlier by differentiating an identity involving Hicksian and Mar-
shallian demands. We start by writing the following arithmetic identity:
xi (P + A P ,m ) - xi ( p ,m ) = xi ( p+ A p , m + Am) - xi ( p ,m )
- [xi( p+ A p , m + Am) - xi ( p+ A p , m ) ] .
Note that this is true by the ordinary rule of algebra.
Suppose that A p = (0,. . . , Apj,. . . , 0 ) . Then the compensating change
in i n c o m e i n the Slutsky sense-is Am = x j ( p ,m)Apj. If we divide each
side of the above identity by Apj and use the fact that Apj = A m / x j ( p ,m ) ,
we have
z i ( p + A P , ~- )x i ( p , m ) - + +
- xi(p A p , m Am) - xi(p,m)
AP~ AP~

-xj(~7m)
+ + +
[ ~ i ( pA p , m Am) - xi(p A p , m ) ]
Am
Interpreting each of the terms in this expression, we can write
Axi
- x.-.
comp 'Am
Note that this last equation is simply a discrete analog of the Slutsky equa-
tion. The term on the left-hand side is how the demand for good i changes
as price j changes. This is decomposed into the substitution effect-how
the demand for good i changes when price j changes and income is also
changed so as t o keep the original level of consumption p o s s i b l e a n d the
income effect-how the demand for good i changes when prices are held
constant but income changes times the demand for good j . The Slutsky
decomposition of a price change is illustrated in Figure 8.7.