Aggregation: Theory

Werner Hildenbrand

University of Bonn
Abstract:
The aim of the aggregation theory is to link the micro and macroeconomic
notions of aggregate demand. One would like such a link to exist for any
heterogeneous population, for a large set of all conceivable income assign-
ments, and for a small number of statistics of the income distribution. This
cannot be achieved. What can be achieved is critically discussed in Section
2. In Section 3, another important topic of aggregation theory is consid-
ered: how does mean demand react to price changes? As an example, the
law of demand is discussed.
1. Introduction
Aggregation theory of demand aims at identifying observable explanatory
variables for aggregate demand starting from a microeconomic descrip-
tion of the underlying population of households. In the simple case, where
the demand decision of a household is the choice of a commodity vector
in a budget set, which is determined by the price vector p and income x
(total expenditure), the demand behaviour of a household h is modeled
by a demand function f
h
(p, x) R
l
+
(commodity space), which is dened
for every strictly positive price vector p P and every income level x 0.
The demand function f
h
might, but need not be derived from preference
maximization under the budget constraint.
Aggregate demand is dened as mean demand across the population H,
that is to say,
1
#H

hH
f
h
(p, x
h
). The population H is viewed as hetero-

Entry for The New Palgrave Dictionary of Economics, 2

nd
Edition.

Forschungsgruppe Hildenbrand, University of Bonn, Lennestr. 37, D-53113 Bonn,

Germany. E-mail: fghildenbrand@uni-bonn.de.
1
geneous in income and demand behaviour. Thus, mean demand is deter-
mined by the price vector p and the joint distribution of income x
h
and
demand function f
h
across the population H.
This general microeconomic denition of mean demand is sufciently
specic for certain problems in pure theory, e.g., for the existence problem
in general equilibrium theory.
In macroeconomics or in applied demand analysis the notion of aggre-
gate demand is quite different. There the explanatory variables for aggre-
gate demand are the price vector and certain statistics S(G
x
) of the income
distribution function G
x
such as mean income, a measure of income in-
equality (e.g., the variance of log income) or higher moments of the income
distribution. In any case, no household specic variable is used in the ag-
gregate demand function. The aim of the aggregation theory is to link the
micro and macroeconomic notions of aggregate demand. More speci-
cally, given an assignment (f
h
)
hH
of demand functions and a set X R
H
+
of income assignments (x
h
)
hH
, one seeks for a representation of mean de-
mand of the following form: there exists a function F from P R
N
into
R
l
+
and N statistics S
1
(G
x
), . . . , S
N
(G
x
) of the income distribution function
G
x
, such that
1
#H

hH
f
h
(p, x
h
) = F(p, S
1
(G
x
), . . . , S
N
(G
x
)) (1)
for all income assignments (x
h
)
hH
in X and all price vectors p in P.
One would like such a representation to exist for any heterogeneous
population H, for a large set X, ideally for all conceivable income assign-
ments, i.e., X = R
H
+
and for a small number N of statistics. This, of course,
cannot be achieved.
The theory of income aggregation is surveyed in section I, where also
basic references are given. The main results are:
a representation of the form (1), which must hold in the case X = R
H
+
is an unreasonable strong requirement. Indeed, if a representation
exists, then the population H must be homogeneous in demand be-
haviour, i.e., f
h
= f for all h H, and furthermore
2
if N is less than the number of households in H and the common
demand function f has the basic properties of demand theory (bud-
get identity and homogeneity), then either f is linear in income or
at least for one commodity i, the income share function w
i
(p, x) :=
p
i
f
i
(p, x)/x is oscillating (i.e., the derivative
x
w
i
(p, ) changes its sign
innitely often).
Thus, households behaviour which is modeled by the common demand
function is either unreasonably simple or incredibly sophisticated. These
results clearly show that the requirement X = R
H
+
leads to an ill-posed
problem.
For a heterogeneous population H there exists (see example 3) a nite
partition {X
k
}
kK
of the set R
H
+
of all conceivable income assignments and
for every k K there is a function F
k
(p, G), where G denotes an income
distribution function, such that
1
#H

hH
f
h
(p, x
h
) = F
k
(p, G
x
) (2)
for every income assignment (x
h
)
hH
in the set X
k
and for every p P.
Thus, for a heterogeneous population H, there is no closed-form def-
inition of an aggregate demand function; there is only a piecewise one,
since the aggregate demand functions F
k
and F
j
are different for k = j.
The less heterogeneous the population the coarser the partition, i.e., the
smaller is #K. The sets X
k
of the partition are large (see Example 3), in
particular, if (x
h
0
) X
k
, then for every strictly increasing function the
income assignment x
h
= (x
h
0
), h H, also belongs to X
k
(see Figures 3
and 4).
The aggregate demand functions F
k
(p, G) in (2) require the knowledge
of the entire income distribution. In many applications one might assume
that the distribution of relevant income assignments in the set X
k
can be
modeled by some few parameters (structural stability of income distribu-
tions). For example, if the population is very large one might restrict
attention to those (x
h
) in X
k
whose distributions are (approximately) log
normal. Then, on this subset of X
k
, mean demand has a representation of
the form F
k
(p, x, ), where x denotes mean income across the population
3
and
2
is the variance of log income, which can be interpreted as a mea-
sure of income inequality.
Another important topic of aggregation theory is to analyse how mean
demand of a heterogeneous population reacts to price changes under the
ceteris paribus clause that households income and demand functions re-
main xed. In this case mean demand is denoted by F(p). Among the
various desirable dependence structures is certainly the law of demand,
which asserts that the vector p R
l
of price changes and the resulting
vector F R
l
of mean demand changes point in opposite directions, i.e.,
the scalar product p F :=

l
i=1
p
i
F
i
is negative.
Certainly, the law is not meant to be an empirical law, but a mono-
tonicity property of the mean demand function F(p) which is dened un-
der a ceteris paribus clause in a mathematical model of a population of
households. Thus, the law asserts that the mean demand function F is
strictly monotone, i.e.,
(p q) (F(p) F(q)) < 0 for all p = q in P.
In particular, every partial mean demand curve is strictly decreasing. This
partial monotonicity property, however, is not sufcient for proving the
uniqueness and stability of the equilibria for a multi-commodity demand-
supply system; one needs strict monotonicity in the multi-commodity ver-
sion.
Which behavioural assumption on the household level and/or which
form of heterogeneity of the population lead to monotone mean demand?
To answer this question one assumes that demand functions f
h
satisfy the
weak axiom of revealed preferences or, more specically, that they are de-
rived from preference maximization. Then, partial monotonicity is eas-
ily obtained, for example, by excluding inferior goods. However, multi-
commodity monotonicity is more difcult to obtain. Trivially, mean de-
mand is monotone if all demand functions f
h
(p, x
h
) were monotone in p.
This, however, requires that either f
h
(p, ) is linear in income or that the
Slutzky substitution effect is sufciently strong. (For a precise formula-
tion, see the Theorem of Mitjuschin and Polterovich, 1978; LAW OF DE-
MAND.) Since the Slutzky substitution effect might be arbitrarily small,
4
one is interested in nding alternative assumptions, which do not rely
on a strong Slutzky substitution effect. These assumptions should not re-
quire that households demand functions are monotone. Obviously, to
obtain the desirable aggregation effect, the population must be hetero-
geneous. Thus, in contrast to the problem of income aggregation, het-
erogeneity does not complicate the analysis, yet it is necessary to obtain
monotonicity of mean demand by aggregation. More details are given in
section II. For example, let H be a population which is homogeneous in
demand behaviour, i.e., f
h
= f, h H and the common demand function
is not monotone. However, the population is heterogeneous in income.
Then, for a given income assignment (x
h
)
hH
, mean demand F
H
(p) is not
monotone in p. If one increases now the population size, i.e., the number
#H of households tends to innity and if for increasing #H the income
distribution functions G
H
of households in H converge to a concave distri-
bution function G, then, for #H sufciently large, mean demand F
H
(p) is
approximately monotone, that is to say, F
H
(p) converges to a monotone
function. Consequently, in the limit, i.e., for an indenitely large popula-
tion which admits a concave income distribution function, mean demand
is monotone. The mathematical model for such a limit population cannot
be a nite or countably innite set; it must be an atomless measure space
of households, e.g., the unit interval [0, 1] with Lebesgue measure (contin-
uum of households).
If these large populations are heterogeneous in income and demand
behaviour, then one can meaningfully pose the problem of smoothing by
aggregation: is mean demand continuous or differentiable without as-
suming these properties on the household level? The basic reference is
Trockel (1984).
Finally, one should mention the literature on behavioural heterogene-
ity initiated by Grandmont (1992). Here the goal is to obtain a stronger
property than strict monotonicity of mean demand: diagonal dominance
of the Jacobian
p
F(p) of mean demand in the sense that
p
i

p
i
F
i
(p) >

j=i
p
j
|
p
j
F
i
(p)|
5
and
p
i

p
i
F
i
(p) >

j=i
p
j
|
p
i
F
j
(p)|.
This diagonal dominance models a strong restriction on the interdepen-
dence among the various commodity markets and is the basis for partial
equilibrium analysis. For a general discussion of behavioural heterogene-
ity see Hildenbrand and Kneip (2005).
2. Income Aggregation
The demand behaviour of every household h in a population H is mod-
eled by a demand function f
h
F. In this section it is not required that
demand functions are derived by preference maximization under budget
constraints. One only needs that demand functions f F are continuous
functions from P R
+
into R
l
+
with f(p, 0) = 0, where P denotes the set
of all strictly positive price vectors in R
l
.
For every income assignment (x
h
)
hH
, x
h
0, we consider mean de-
mand
1
#H

hH
f
h
(p, x
h
). The problem of income aggregation has been
dened in the literature by the question: does there exist a function F
from P R
+
into R
l
+
such that
1
#H

hH
f
h
(p, x
h
) = F(p, x), where x =
1
#H

H
x
h
, (3)
for all income assignments in a given set X R
H
+
and all p P?
If one asks this question for all conceivable income assignments, i.e.,
X = R
H
+
, then this is an ill-posed problem since it allows only a trivial so-
lution.
Theorem (Antonelli, 1886): There exists a function F(p, x) such that (1) holds
on R
H
+
P if and only if the population H is homogeneous in demand behaviour,
i.e., f
h
= f, and furthermore f(p, x) is linear in x, i.e., f(p, x) = (p)x,
(p) R
l
+
. Thus, F(p, x) = (p) x.
One might ask whether a less restrictive condition than (3) allows for a
nontrivial solution. That is to say, one might consider mean demand func-
6
tions that depend on a wider set of aggregate income variables than just
mean income, for example, the variance or higher moments of the distri-
bution of income. The answer is denitely negative.
For every income assignment (x
h
)
hH
, let G
x
denote its distribution
function, i.e.,
G
x
() :=
1
#H
#{h H | x
h
}, R.
7
Proposition 1. There exists a function F(p, G) such that
1
#H

hH
f
h
(p, x
h
) = F(p, G
x
) (4)
for all conceivable income assignments, i.e., X = R
H
+
and all p P, if and only if
the population H is homogeneous in demand behaviour, i.e., all households in H
have the same demand function. Then F(p, G
x
) =
_
f(p, )dG
x
().
Proof: Consider any two households k and j in H, and an income assign-
ment (x
h
)
hH
with x
k
> 0 and x
j
= 0. Now one interchanges the income
of households k and j. This does not change the distribution function of
income. Hence property (4) and the fact that f
k
(p, 0) = f
j
(p, 0) = 0 im-
plies that f
k
(p, x
k
) = f
j
(p, x
k
). Since this holds for all x
k
> 0 and p P
one obtains f
k
= f
j
. On the other hand, if f
h
= f for all h H, then
1
#H

hH
f
h
(p, x
h
) =
_
f(p, x)dG
x
=: F(p, G
x
).
The justication for considering the generalized problem of income ag-
gregation as dened by (4) is based on the view that for large populations,
which this survey emphasizes, income distribution functions can often be
modeled by some few parameters, e.g. log-normal distributions.
By Proposition 1 it is clear that one is forced to restrict the set X of
admissible income assignments if one wants to escape the case of trivial
solutions, f
h
= f, to the aggregation problem as dened by (4). Motivated
by the special role which zero income and the assumption f(p, 0) = 0 play
in the proofs of Antonellis Theorem or Proposition 1 one has considered
in the literature (e.g. Nataf, 1948 or Gorman, 1953) a restriction on the
domain of individual income:
X(a, b) := {(x
h
) R
H
+
| 0 < a x
h
b }, a < b.
Proposition 2 shows that this restriction allows merely for some very
limited and quite special heterogeneity in demand behaviour of the pop-
ulation H.
8
Proposition 2.
(i) There exists a function F(p, G) such that (2) holds on X(a, b) P if and
only if for every commodity i and p P the income expansion paths
f
h
i
(p, ), h H, are parallel (vertically) on the interval (a,b); (with dif-
ferentiability)
x
f
h
i
(p, x) does not depend on h H (Figure 1).
(ii) There exists a function F(p, x) such that (1) holds on X(a, b) P if and
only if for every commodity i and p P the income expansion paths
f
h
i
(p, ), h H, are afne and parallel on the interval (a,b); (with differ-
entiability)
x
f
h
i
(p, x) does not depend on h H and x (a, b) (Figure
2).
(iii) If all individual demand functions f
h
belong to F and are homogeneous in
(p, x), then the necessary condition in (i) implies that f
h
f.
(Insert Figure 1 about here)
(Insert Figure 2 about here)
Proof:
(i) Consider any two households k and j in H and an income assignment
in X(a, b) with x
k
= x
j
. Now one interchanges the income of house-
holds k and j. This does not change the income distribution function.
Hence, property (2) implies f
k
(p, x
k
)+f
j
(p, x
j
) = f
k
(p, x
j
)+f
j
(p, x
k
). Thus
f
k
(p, x
k
)f
k
(p, x
j
) = f
j
(p, x
k
)f
j
(p, x
j
). Since it holds for all x
k
, x
j
(a, b)
and all p P one obtains the claimed property in (i). The converse is triv-
ial.
(ii) Instead of interchanging the income of households k and j one chooses
x
k
+ and x
j
(a, b) for sufciently small . Property (1) then implies
f
k
(p, x
k
+)f
k
(p, x
k
) = f
j
(p, x
j
)f
j
(p, x
j
) = f
k
(p, x
j
)f
k
(p, x
j
)
by (i), which implies the claimed property in (ii). The converse is trivial.
(iii) If the expansion paths f
k
i
(p, ), h H, are parallel on (a, b) for ev-
ery p P, then homogeneity implies that they are also parallel on (a, b)
9
for all > 0 and p P. Hence they are parallel on (0, ) for all p P.
Continuity and f
h
(p, 0) = 0 then implies the claim.
An alternative approach to allow for heterogeneous populations con-
sists of considering, in addition to income, further explanatory variables
for household demand. For example, in applications it is standard practice
to stratify the whole population H by a certain prole a = (a
1
, a
2
, . . . ) of
observable household attributes, such as household size, age of household
head, etc. Let H(a) denote the subpopulation of all households in H with
attribute prole a. Without loss of generality one can assume that a R
m
.
Let G
x,a
denote the joint distribution of function of x
h
, a
h
across H. Analo-
gously to Proposition 1 one shows
Proposition 1

. There exists a function F(p, G

x,a
) such that
1
#H

hH
f
h
(p, x
h
) = F(p, G
x,a
)
for all conceivable income-attribute assignments and all p P if and only if all
subpopulations H(a) are homogeneous in demand behaviour, i.e., f
h
= f
a
for all
h H(a).
Thus, the whole population need not be homogeneous, yet the joint
distribution of x
h
and a
h
across H has typically a complex dependence
structure, and hence, it cannot be modeled by some few parameters, as in
the case of income.
Exact income aggregation
In the literature on exact income aggregation, as initiated by Gorman
(1953), Lau (1982), and Jorgenson et al. (1982), one seeks for a representa-
tion of mean demand which is less restrictive than (3), yet more demand-
ing than (4), that is to say,
1
#H

hH
f
h
(p, x
h
) = F(p, S
1
(G
x
), . . . , S
N
(G
x
))
on R
H
+
P for some continuous function F from P R
N
into R
l
(the com-
modity space) and some vector of distributional statistics S
1
(G
x
), . . . , S
N
(G
x
)
with N < #H. This representation is more demanding than (4); it does not
require the knowledge of the entire income distribution since N < #H.
10
If such a representation exists, then by Proposition 1, f
h
= f, h H,
and f is called exactly aggregable. Thus, the question is whether there
are exactly aggregable demand functions which are not linear in income
and satisfy the basic restrictions of demand theory?
To simplify the presentation one assumes that all distributional statis-
tics are generalized moments, i.e., S
n
(G
x
) =
_
s
n
()dG
x
(), with con-
tinuous functions s
n
(). Without loss of generality one can require that
s
n
(0) = 0.
Proposition 3. There exists a representation of mean demand of the form
_
f(p, )dG
x
() = F
_
p,
_
s
1
()dG
x
(), . . . ,
_
s
N
()dG
x
()
_
, (5)
which holds for every income distribution function G
x
of every nite population
H and every price vector in P if and only if the function f is of the form
f(p, ) =
1
(p)s
1
() +
N
(p)s
N
(), p P and R
+
, (6)
where
n
(p) R
l
.
Proof: Trivially, (6) implies (5). Assume that (5) holds. Let G denote the
set of all income distribution functions for every nite population. Note
that for every G
1
, G
2
G and any rational with 0 1 it follows that
G

= G
1
+ (1 )G
2
G.
The representation (5) implies for every commodity i
f
i
(p, ) = F
i
_
p, s
1
(), . . . , s
N
()
_
, p P and R
+
. (7)
Now one shows that the function F
i
(p, ) has a linear structure on its rel-
evant domain D := {y R
N
| y
n
=
_
s
n
()dG(), G G, n = 1, . . . , N},
i.e.,
F
i
(p, y
1
+ (1 )y
2
) = F
i
(p, y
1
) + (1 )F
i
(p, y
2
) (8)
for every y
1
, y
2
D and any rational with 0 1. Indeed, y
k
n
=
_
s
n
()dG
k
(), k = 1, 2 for some G
1
, G
2
G. Let G

= G
1
+ (1 )G
2
.
Then
_
s
n
()dG

() =
_
s
n
()dG
1
() +(1 )
_
s
n
()dG
2
(). Hence y
1
+
11
(1 )y
2
D since G

G for rational . Consequently, the closure

D of
D is convex. Since G

G, one obtains from (5)

_
f
i
(p, )dG

() = F
i
_
p,
_
s
1
()dG

(), . . . ,
_
s
N
()dG

()
_
= F
i
(p, y
1
+(1)y
2
).
The left hand side is equal to
_
f
i
(p, )dG
1
() + (1 )
_
f
i
(p, )dG
2
() =
F
i
(p, y
1
) + (1 )F
i
(p, y
2
) by (5), which proves (8). Since F
i
is continu-
ous, the linear structure (8) also holds for any y
1
, y
2
in the closure

D of
D and any with 0 1. Since s
n
(0) = 0 and f(p, 0) = 0 it follows
from (7) that F
i
(p, 0) = 0. Consequently, by (8), the restriction of the func-
tion F
i
(p, ) on the convex domain

D can be extended to a function

F
i
(p, ),
which is linear in y, i.e.,

F
i
(p, y) =
i
1
(p)y
1
+
i
N
(p)y
N
. Thus (7) implies
(6). The extension is unique if the dimension of the convex domain

D is
equal to N.
Remark: The proof of Proposition 3 is quite simple since it was assumed
that the representation (5) must hold for all income distribution functions
for all nite populations. This case is also treated in Heineke and Shefrin
(1988), their proof, however, requires differentiability. If one only requires
(5) to hold for all income distribution functions of a given population H
with N < #H, then it is much more difcult to obtain (6). See Lau (1982)
and Heineke and Shefrin (1988).
Note that the global structural specication (6) is very restrictive if the
demand function f F has the basic properties of static demand theory.
In fact, Heineke and Shefrin (1987) show the following result: if f F sat-
ises the budget-identity, is homogeneous in p and x and if no budget share func-
tion w
i
(p, x) := p
i
f
i
(p, x)/x is oscillating (i.e., the derivative
x
w
i
(p, x) changes
innitely often its sign), then (6) implies f(p, x) = (p)x.
Indeed, if f F satises the budget identity, then 0 w
i
(p, x) 1.
Let the budget share function w
k
( p, ) be non-constant and non-oscillating.
Consider the function

(), > 0, dened by

(x) = w
k
( p, x), and
the linear function space which is generated by all functions

(), > 0.
Heineke and Shefrin (1987) argue that the dimension of this linear space
is innite. By homogeneity,

(x) = w
k
( p/, x); thus, the linear space L
which is generated by all budget share functions w
k
(p, ), p P has in-
nite dimension. Consequently, the demand function f cannot satisfy (6),
12
since (6) implies that dimL N. Thus, if f satises (6) and w
k
( p, ) is non-
oscillating, then it must be constant, i.e., f
k
( p, ) is linear.
As a consequence, for demand functions which have the basic proper-
ties of atemporal demand theory including non-oscillating budget share
functions, one either has to be satised with a representation as in Propo-
sition 1 or one is in the trivial case of Antonellis Theorem.
Heterogeneous populations
The representations (3), (4), and (5) of mean demand which have been con-
sidered up to now imply that the population of households must be ho-
mogeneous in demand behaviour, i.e., f
h
f, h H. The reason for this
unsatisfactory fact is due to the very strong requirement that the represen-
tations must hold for every conceivable income assignment. This is more
demanding than is needed in many applications, since there, changes in
individual income are not entirely arbitrary; they might be the result of an
underlying process. This point was emphasized by Malinvaud (1956) and
(1993). To capture this idea, one starts from an initial income assignment
(x
h
0
) (status quo), and then one considers a sequence (x
h
n
), n = 1, 2, . . . or a
set X(x
0
) of income assignments which are viewed as the result of the un-
derlying (unspecied) process. Which properties must the sequence (x
h
n
)
or the set X(x
0
) have such that for any assignment of demand functions
f
h
the representations of mean demand hold along this sequence or on the
set X(x
0
)?
We give three examples. The rst one is well-known. The second and
third example generalize substantially the rst one.
Example 1: Fixed income shares
Starting froman initial income assignment (x
h
0
), one denes the set X() R
H
+
of income assignments
X() :=
_
(x
h
) R
H
+
| x
h
/ x = x
h
0
/ x
0
=:
h
_
.
where x denotes mean income across H.
13
Given any assignment of demand functions f
h
, h H, there exists a function
F from P R
+
into R
l
+
such that mean demand has the representation
1
#H

hH
f
h
(p, x
h
) F(p, x) on X() P. (9)
The function F is dened by F(p, x) =
1
#H

hH
f
h
(p,
h
x). If all f
h
are
linear in income then F(p, x) is linear in mean income x. Moreover, Eisen-
berg (1961) and Chipman and Moore (1979) have shown: if all f
h
are gen-
erated by a utility function homogeneous of degree one then F(p, x) is also
generated by a utility function homogeneous of degree one given by
u(z) = max
z
h
R
l
+
,

H
z
h
=z

hH
(u
h
(z
h
))
x
h
0
/ x
0
Example 2: Rank preserving income changes.
Starting from an initial income assignment (x
h
0
)
hH
one denes the set
X(x
0
) R
H
+
of income assignments (x
h
) which have the property that
every household keeps his rank position of income, i.e., if for two house-
holds j and k, x
j
0
= x
k
0
then x
j
= x
k
and if x
j
0
< x
k
0
then x
j
< x
k
. For any
(x
h
1
) and (x
h
2
) in X(x
0
) there is a strictly increasing function such that
(x
h
1
) = x
h
2
, h H. Examples for () are given in Figure 3 (low income is
increased, high income is decreased) and Figure 4 (low and high incomes
are decreased, middle ones increased) below.
(Insert Figure 3 about here)
(Insert Figure 4 about here)
Note that (x
h
) X(x
0
) implies X(x) = X(x
0
) and (x
h
) / X(x
0
) implies
X(x) X(x
0
) = . Thus, there is a nite partition {

X
i
} of R
H
+
into sets

X
i
of rank preserving income assignments.
Note that for any rank preserving income assignments (x
h
) in X(x
0
)
one can recover the income assignment from knowing only its distribution
function G
x
, since x
h
= G
1
x
G
0
(x
h
0
) for every h H, where G
1
denotes the
quantile function (quasi-inverse) of the distribution function G, which is
14
dened by G
1
(q) := inf{x R
+
| G(x) q}. Consequently, one obtains:
Given any assignment of demand functions f
h
, h H, there exists a function
F(p, G) such that mean demand has the representation
1
#H

hH
f
h
(p, x
h
) F(p, G
x
) on X(x
0
) P. (10)
The function F is dened by F(p, G
x
) =
1
#H

hH
f
h
(p, G
1
x
G
0
(x
h
0
)).
There might be larger sets than X(x
0
) for which the representation (10)
holds. For example, if households k and j have the same demand function
then one can interchange their rank position. Thus, in dening a set X for
which (10) holds, one should take into account the heterogeneity structure
of (f
h
)
hH
. This is done in the next example
Example 3: Common copula
Let {f
1
, . . . , f
N
} be the set of distinct demand functions of the given as-
signment (f
h
)
hH
. Thus, for h H there is an integer n(h) N such that
f
h
= f
n(h)
. For every income assignment (x
h
)
hH
consider the bivariate
distribution function D
x
, which is dened by
D
x
(, ) :=
1
#H
#{h H| x
h
and n(h) }, , R.
The distribution function D
x
and the price vector p determines mean de-
mand
1
#H

hH
f
h
(p, x
h
). The marginal distribution functions of D
x
are
denoted by G
x
and V .
By Sklars Theorem (see, e.g., Nelson, 1999), for every bivariate dis-
tribution function D with marginals G and V , there exists a copula C (a
function from [0, 1]
2
into [0, 1] with certain properties) such that D(, ) =
C(G(), V ()) for all , R. Conversely, if C is a copula and G and V are
distribution functions, then C(G(), V ()) is a bivariate distribution func-
tion. Thus, a copula couples the marginals to the bivariate distribution.
The copula models the dependence structure of the bivariate distribution
function.
15
Starting from an initial income assignment (x
h
0
), one considers the set
X(x
0
, f) R
H
+
of income assignments (x
h
) such that the corresponding
bivariate distribution functions D
x
have a common copula. Thus, the de-
pendence structure of (x
h
, f
h
) across H is the same for all (x
h
) in X(x
0
, f).
It follows that income assignments in the set X(x
0
) of rank preserving in-
come assignments is contained in the set X(x
0
, f). Furthermore,
given any assignment of demand functions (f
h
)
hH
, there exists a function F(p, G)
such that mean demand has the representation
1
#H

hH
f
h
(p, x
h
) F(p, G
x
) on X(x
0
, f) P.
There is a very simple, however, special case which is worthwhile to
be mentioned (and could have been discussed at the beginning). If the
initial income x
h
0
and the demand function f
h
of household h are indepen-
dently distributed across H, i.e., D
x
0
(, ) G
x
0
()V () (the copula of D
xo
is equal to C(u, v) = u v), then the set X(x
0
, f) =: Z(x
0
) is very large; it
consists of all income assignments (x
h
) R
H
+
with the property: x
k
0
= x
j
0
implies x
k
= x
j
. Then, one obtains
1
#H

hH
f
h
(p, x
h
) F(p, G
x
) on Z(x
0
) P
with F(p, G) =
_

f(p, )dG() where

f(p, ) =
1
#H

hH
f
h
(p, ).
3. Monotone mean demand
The law of demand for a population of households asserts that the vec-
tor of price changes p R
l
and the resulting vector of mean demand
changes F R
l
point in opposite directions, provided the price changes
do not affect households incomes (total expenditure) and demand func-
tions (preferences). Thus, the law asserts that the mean demand function
F(p) is strictly monotone, i.e.
(p q) (F(p) F(q)) < 0 for all p, q R
l
++
, p = q
Strict monotonicity of mean demand implies, in particular, that for ev-
ery commodity i the partial mean demand function F
i
is strictly decreas-
ing in its own price p
i
and that the mean demand function F() is invertible
16
(existence of an inverse demand function).
The goal of aggregation theory is to derive strict monotonicity of mean
demand without assuming that households demand functions f
h
(p, x) are
strictly monotone in p.
Demand functions f
h
F are assumed to be continuous in p and x
and satisfy the budget-identity p f(p, x) = x. The function f F satises
the Weak Axiom of revealed preferences if for every price-income pair (p, x)
and (p

, x

), p f(p

, x

) x implies p f(p, x) x

, and satises the Axiom

of revealed preferences, if f(p, x) = f(p

, x

) and p f(p

, x

) x implies
p

f(p, x) > x

.
Every demand function which is derived from a continuous, strictly
convex and non-saturated preference relation satises the Axiom, yet it is
not necessarily monotone.
Theorem (Hildenbrand, 1983)
1. The function F(p) :=
_

0
f(p, x)(x)dx is monotone, i.e.,
(p q) (F(p) F(q)) 0 for all p, q in R
l
++
, if f F satises
the Weak Axiom of revealed preferences and is a density which is
non-increasing on R
+
with
_

0
(x)dx <
2. The function F is strictly monotone, if, in addition, f satises the
Axiom of revealed preferences and the expansion paths f(p, .) and
f(q, .) have only 0 in common for any p, q that are not collinear.
Interpretation: The underlying micro-model is a population H of house-
holds which is indenitely large; mathematically, an atomless measure
space, e.g. the unit interval [0, 1] with Lebesgue measure. Every household
h [0, 1] is modeled by its income x(h) 0 and the common demand func-
tion f. The income assignment x() is an integrable function whose distri-
bution admits a density . Thus, mean demand F(p) =
_
1
0
f(p, x(h))dh =
_

0
f(p, x)(x)dx.
Three questions are relevant:
1. Why a continuum of households? Does the result still hold approxi-
mately for a large but nite population?
17
2. Why a non-increasing income density? Does monotonicity of F fail
if the density is rst increasing and then decreasing?
3. Why a common demand function? Does the result extend to hetero-
geneous populations in income and demand behaviour?
The discussion of these questions is simplied by assuming that f is
continuously differentiable in p and x. Then monotonicity of F is equiva-
lent with negative semi-deniteness (n.s.d.) of the Jacobian matrix
p
F(p)
for all p, i.e.,

l
i,j=1
v
i
v
j

p
i
F
j
(p) 0 for all v R
l
, and the Weak Axiom
for f is equivalent with n.s.d. of the Slutzky substitution matrix. Con-
sequently, monotonicity of F follows from the positive semi-deniteness
(p.s.d.) of the mean income effect matrix I(f, ) =
_
I(f, x)(x)dx, where
I(f, x) = (f
i
(p, x)
x
f
j
(p, x))
i,j=1,...,l
.
Question 1: The mean income effect matrix for a nite population H,
i.e.,
1
#H

H
(f
i
(p, x
h
)
x
f
j
(p, x
h
))
i,j
= I
H
is p.s.d. if and only if for ev-
ery v R
l
, v I
H
v =
1
#H

H
g

(x
h
) 0 where g(x) :=
1
2
(v f(p, x))
2
.
Assume that income x
h
is measured in multiples of (euro). Let
n
:=
1
#H
#{h H | x
h
= n =: x
n
}, n = 0, 1, . . . Then
(1)
1
#H

H
g

(x
h
) =

n=0

n
g

(x
n
) =

n=1
1

(
n1

n
)g(x
n
) + o()
using the approximation
(2) g

(x
n
) =
1

(g(x
n+1
) g(x
n
)) + o().
Consequently, one needs
n1

n
, n = 1, ..., to obtain a non-negative
rst term on the right hand side of (1); this is the nite analogue of a
non-increasing density. Thus, for a nite population with a small (which
requires by
n1

n
a large population) one obtains the desired result
up to the small term o(). For a population H = [0, 1] one does not need
the approximation (2) and hence o(), since (1) becomes
_
g

(x)(x)dx =

_
g(x)

(x)dx (by partial integration), which is non-negative for a non-

increasing differentiable density .
Question 2: The mean income effect matrix I(f, ) is p.s.d. in each of the
two extreme cases: either, is non-increasing and no assumption on the
18
shape of the income expansion path f
i
(p, ) or, no assumption on yet lin-
earity of f
i
(p, ). There must be results in between. Indeed, if the curvature
of all income expansion paths f
i
(p, ) is limited and the unimodal density
is sufciently skewed, then I(f, ) is p.s.d.
Example: All income expansion paths restricted to the interval [0, x] are
polynomials of degree n (note that, no non-linear f
i
(p, ) can be a poly-
nomial on R
+
) and is concentrated on [0, x]. Then, I(f, g) is p.s.d. if
and only if the matrix M(n, ) := ((i + j)m
i+j1
)
i,j=1,...,n
is p.s.d. where
m
k
:=
_
x
k
(x)dx (Hildenbrand, 1994, Appendix 6).
Let the densities
m
be as in Figure 5.
< Figure 5 here >
For every n there exists m(n) > 0 such that I(f,
m
) is p.s.d. if m m(n);
e.g. n = 2, m(2) = 0.38 x or n = 3, m(3) = 0.14 x.
For a more general analysis see Chiappori (1985) and Hildenbrand
(1994).
Question 3: A population of households that is heterogeneous in income
and demand functions is described by a joint distribution of income and
demand functions, i.e., is a distribution on R
+
F. (Areader not familiar
with distributions on function spaces might replace F by a nite set F
0
).
As before, the marginal distribution of income admits a density . The
conditional distribution of demand functions given the income level x is
denoted by (x). Then mean demand
F(p) :=
_
R
+
F
f(p, x)d =
_

0

f(p, x)(x)dx
where

f(p, x) :=
_
F
f(p, x)d(x). Consequently, the Theorem or the exten-
sions discussed under Question 2 imply that F(p) is monotone provided
the function

f satises the Weak Axiom. This approach to derive mono-
tonicity for a heterogeneous population is the most direct, yet not the most
general way (see Hildenbrand, 1994).
19
It is well-known (Hicks, 1956, p.53) that

f does not necessarily satisfy
the Weak Axiom, even if individual demand functions are derived from
utility maximization. The following two assumptions (which, again, are
not the most general ones) imply that

f satises the Weak Axiom
(a) independence: (x) does not depend on x
(b) increasing dispersion: the distribution D(x + ), > 0, is more dis-
persed than the distribution D(x), where D() denotes the distribu-
tion (in the commodity space R
l
) of individual demand of all house-
holds with income at the price p (i.e., D() is the image distribution
of under the mapping f f(p, )).
Generalizing the one-dimensional case where the variance is a mea-
sure of dispersion one chooses the positive deniteness of the covariance
matrix as a measure of dispersion for distributions on R
l
. Thus, increasing
dispersion means that for > 0, covD(x + ) covD(x) is positive semi-
denite.
Assumptions (a) and (b) are quite restrictive, in particular, the inde-
pendence assumption. Therefore one partitions the whole population H
into subpopulations H(a) by stratifying with respect to a certain vector a
of household attributes (household size, age, ...) and then one requires as-
sumptions (a) and (b) for each subpopulation H(a). The role of stratifying
is to reduce the heterogeneity in demand behaviour. In the extreme case,
where stratifying leads to a homogeneous subpopulation in demand be-
haviour, assumptions (a) and (b) are trivially satised. If the income den-
sity of each subpopulation H(a) is non-increasing on R
+
or if the extension
discussed in Question 2 apply, the mean demand of each subpopulation
is monotone and hence also the mean demand of the whole population,
since monotonicity is additive.
A more general denition of increasing dispersion and a detailed dis-
cussion is given in Hildenbrand (1994). For an empirical study of the Law
of Demand see H ardle et al. (1991).
A broader discussion of the law of demand and related properties in-
cluding cases where income is price dependent is contained in the entry
20
LAW OF DEMAND.
21
Bibliography
Antonelli, G.B. 1886. Sulla teoria matematica della economia politica. Pisa:
Nella Tipograa del Folchetto (translated by J.S. Chipman and A.P. Kir-
man in Preferences, utility and demand, ed. J.S. Chipman et al. New
York: Harcourt Brace Jovanovich, 33360.)
Chiappori, P. 1985. Distribution of income and the law of demand. Econo-
metrica 53, 109127.
Chipman, J.S. and Moore, J. 1979. On social welfare functions and the ag-
gregation of preferences. Journal of Economic Theory 21, 11139.
Eisenberg, B. 1961. Aggregation of utility functions. Management Science
7, 33750.
Gorman, W.M. 1953. Community preference elds. Econometrica 21, 63
80.
Grandmont, J.M. 1992. Transformations of the commodity space, behavioural
heterogeneity, and the aggregation problem. Journal of Economic Theory
57(1), 1-35.
H ardle, W., Hildenbrand, W. and Jerison, M. 1991. Empirical evidence on
the law of demand. Econometrica 59, 15251549.
Heineke, J. and Shefrin, H. 1987. On some global properties of Gorman
class demand systems. Economics Letters 25, 155160.
Heineke, J. and Shefrin, H. 1988. Exact aggregation and the nite basis
property. International Economic Review 29(3), 525538.
Hicks, J.R. 1956. A Revision of Demand Theory. London: Oxford Univer-
sity Press.
Hildenbrand, W. 1983. On the lawof Demand. Econometrica 51, 9971019.
Hildenbrand, W. 1994. Market Demand. Princeton: Princeton University
22
Press.
Hildenbrand, W. and Kneip, A. (2005). On behavioral heterogeneity. Eco-
nomic Theory 25, 155169.
Jorgensen, D.W., Lau, L. and Stoker, T.M. 1982. The transcendental loga-
rithmic model of aggregate consumer behavior. (in: Basmann and Rhodes
eds. Advances in Econometrics. Greenwich: JAI Press), 97238.
Lau, L. 1982. A note on the fundamental theorem of exact aggregation.
Economics Letters 9, 119126.
Malinvaud, E. 1956. Lagregation dans les modeles economiques. Cahiers
du Seminaire dEconometrie 4, 69146.
Malinvaud, E. 1993. A framework for aggregation theories. Ricerche Eco-
nomiche 47(2), 107135.
Mitjuschin, L.G. and Polterovich, W.M. 1978. Criteria for monotonicity of
demand functions. Ekonomika i Matematicheskie Metody 14, 122128. (In
Russian.)
Nataf, A. 1948. Sur la possibilite de construction de certains macromode-
les. Econometrica 16, 232244.
Nelson, R. 1999. An introduction to copulas. Lecture notes in statistics
139. New York: Springer Verlag.
Trockel. W. 1984. Market demand, an analysis of large economies with
non-convex preferences. Lecture notes in economics and mathematical
systems. Vol. 223. Heidelberg: Springer Verlag.
23
-
6
f
h
i
(p, )
income a b
1
Figure 1:
-
6
f
h
i
(p, )
income a b

1
Figure 2:
24
-
6

income
x
h
x
h
0
6
x
j
0
?
strictly increasing function
x
j

1
Figure 3:
-
6

x
l
0
?
x
h
0
x
m
0
6
?
x
l
x
m
x
h

1
Figure 4:
25
-
6

Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
ZZ
x m
2
x

m
-
1
Figure 5:
26