Division of the Humanities

and Social Sciences

Technical Results on Regular Preferences and Demand

KC Border
Revised Fall 2011; Winter 2017
v. 2019.10.30::12.40

Preferences
For the purposes of this note, a preference relation (or simply a preference) ≽ on a set X
is a reflexive, total, transitive binary relation on X. Following Richter [3], I usually call this a
regular preference. Mas-Colell, Whinston, and Green [2] call this a rational preference.
There are times we wish to consider a weaker notion of preference that may be incomplete or
non-transitive, but for now a preference is always reflexive, total, and transitive.1 The binary
relations ≻ and ∼ are the asymmetric and symmetric parts of ≽, defined by

x≻y if x ≽ y and not y ≽ x

and
x∼y if x ≽ y &y ≽ x.
The symmetric part ∼ of a preference relation is called the induced indifference relation, and
is reflexive, transitive, and symmetric. That is, ∼ is an equivalence relation. The equivalence
classes of ∼ are traditionally called indifference curves, even though in general they may not
be “curves.”2 The asymmetric part ≻ of ≽ is called the induced strict preference relation. It
is transitive, irreflexive, and asymmetric.
Recall that a function u : X → R is a utility for ≽ if

x≽y ⇐⇒ u(x) ⩾ u(y).

1
Recall that a binary relation ≽ on X is (i) reflexive if ( ∀x ∈ X ) [ x ≽ x ];
(ii) total if ( ∀x, y ∈ X ) [ x ̸= y =⇒ (x ≽ y or y ≽ x) ]; and (iii) transitive if
( ∀x, y, z ∈ X ) [ (x ≽ y & y ≽ z) =⇒ x ≽ z ]. Some authors use the term complete to mean reflexive
and total, that is, ( ∀x, y ∈ X ) [ x ≽ y or y ≽ x ]. Debreu [1] calls a reflexive and transitive relation a preorder
or quasiorder. Getting ahead of ourselves for a moment, a binary relation ∼ on X is (iv) symmetric if
( ∀x, y ∈ X ) [ x ∼ y =⇒ y ∼ x ]. The relation ≻ is (v) irreflexive if ( ∀x ∈ X ) [ ¬x ≻ x ]; and (vi) asymmetric
if ( ∀x, y ∈ X ) [ x ≻ y =⇒ ¬y ≻ x ].
2
The equivalence class [x] of an element x of X is simply {y ∈ X : y ∼ x}. The equivalence classes partition
X. That is, each x belongs to some equivalence class (namely [x]); and for two equivalence classes [x] and [y]
either [x] = [y] (which occurs when x ∼ y) or [x]∩[y] = ∅. Economists often phrase this latter fact as “indifference
curves do not cross.”

1
KC Border Technical Results on Regular Preferences and Demand 2

Nonsatiation
A preference relation ≽ on a set X has a satiation point x if x is a greatest element, that is,
if x ≽ y for all y ∈ X. A preference relation is nonsatiated if it has no satiation point. That
is, for every x there is some y ∈ X with y ≻ x.
If (X, d) is a metric space, the preference relation ≽ is locally nonsatiated if for every
x ∈ X and every ε > 0, there exists a point y ∈ X with d(y, x) < ε and y ≻ x. Note that this is
a joint condition on X and ≽. In particular, if X is nonempty, it must be that for each point
x ∈ X and every ε > 0 there exists a point y ̸= x belonging to X with d(y, x) < ε. That is, X
cannot have isolated points.

Monotonicity
Now let X be a subset of Rn . We use the following notation for partial orders on Rn :

x≧y if xi ⩾ yi , i = 1, . . . , n;
x>y if x ≧ y & x ̸= y; and
x≫y if xi > yi , i = 1, . . . , n.

A preference ≽ on X is

monotonic if x ≫ y =⇒ x ≻ y, and
strictly monotonic if x > y =⇒ x ≻ y.

The next lemma demonstrates that monotonicity implies local nonsatiation. The converse
is not true, but we shall see in Proposition 23 below that the apparent increase in generality is
illusory.

1 Lemma (Monotonicity implies local nonsatiation for Rn+ ) Let ≽ be a monotonic

preference on Rn+ . Then ≽ is locally nonsatiated.

Proof : Let x belong to Rn+ , let 1 ∈ Rn be the vector whose components are all one, and let
ε > 0 be given. Now for any λ > 0, we have that x + λ1 ∈ Rn+ and x + λ1 ≫ x, so by
√ √
monotonicity, x + λ1 ≻ x. Now d(x + λ1, x) = ∥x + λ1 − x∥ = λ∥1∥ = λ n, so for λ < ε/ n,
the point y = x + λ1 satisfies d(y, x) < ε and y ≻ x. Since x and ε were arbitrary, ≽ is locally
nonsatiated.

Continuity
Given a preference relation ≽ on a set X, define the strict and weak upper contour sets

P (x) = {y ∈ X : y ≻ x} and U (x) = {y ∈ X : y ≽ x}.

We also define the strict and weak lower contour sets

P −1 (x) = {y ∈ X : x ≻ y} and U −1 (x) = {y ∈ X : x ≽ y}.

When (X, d) is a metric space, we say that ≽ is upper semicontinuous if for each x, the
set U (x) is closed, or equivalently (since ≽ is total), P −1 (x) is open in X. Similarly, ≽ is lower

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 3

semicontinuous if for each x, the set U −1 (x) is closed, or equivalently, P (x) is open in X. A
preference relation ≽ is continuous if and only if it is both upper and lower semicontinuous.
The next lemma gives several equivalent characterization of continuity for preferences.

2 Lemma (Continuity of preferences) For a total, transitive, reflexive preference relation

≽ on a metric space X, the following are equivalent.

1. The graph of ≽ is closed. That is, if yn → y, xn → x, and yn ≽ xn for each n, then y ≽ x.

2. The graph of ≻ is open. That is, if y ≻ x, there is an ε > 0 such that if d(y ′ , y) < ε and
d(x′ , x) < ε, then y ′ ≻ x′ .

3. For each x, the weak contour sets U (x) = {y ∈ X : y ≽ x} and U −1 (x) = {y ∈ X : x ≽ y}

are closed.

4. For each x, the strict contour sets P (x) = {y ∈ X : y ≻ x} and P −1 (x) = {y ∈ X : x ≻ y}

are open.

Proof : Since ≽ is total, it is clear that (1) ⇐⇒ (2) and (3) ⇐⇒ (4). Moreover it is also
immediate that (1) =⇒ (3) and (2) =⇒ (4). So it suffices to prove that (4) implies (1).
So assume by way of contradiction that yn → y, xn → x, and yn ≽ xn for each n, but x ≻ y.
Since P (y) is open by condition (4) and x ∈ P (y) by hypothesis, there is some ε > 0 such that
d(z, x) < ε implies z ∈ P (y), or z ≻ y. Similarly, since P −1 (x) is open and y ∈ P −1 (x) there is
some ε′ > 0 such that d(w, y) < ε′ implies x ≻ w. Since xn → x and yn → y, for large enough
n, we have d(xn , x) < ε and d(yn , y) < ε′ , so

x ≻ yn ≽ xn ≻ y

for these large n. Pick one such n, call it n0 , and observe that

x ≻ xn0 ≻ y.

Now condition (4) implies P (xn0 ) is open and since x ∈ P (xn0 ), there is some η > 0 such
that d(z, x) < η implies z ≻ xn0 . Similarly, since P −1 (xn0 ) and y ∈ P −1 (xn0 ), there is η ′ > 0
such that d(w, y) < η ′ implies xn0 ≻ w. Now for large enough n we have d(xn , x) < η and
d(yn , y) < η ′ , so
xn ≻ xn0 ≻ yn ,
which contradicts yn ≽ xn for all n.

The following proposition is almost trivial. The converse is considerably more involved, and
requires some additional assumptions on X (such as second countability and connectedness).

3 Proposition If ≽ has a utility that is a continuous function, then ≽ is a continuous prefer-

ence.

4 Lemma If ≽ is upper semicontinuous and locally nonsatiated, then U (x) is the closure of
P (x).

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 4

Proof : P (x) ⊂ U (x): Let y belong to P (x). That is, there is a sequences yn in P (x) ⊂ U (x)
with yn → y. Now U (x) is closed by upper semicontinuity, so y ∈ U (x).
U (x) ⊂ P (x): Let y belong to U (x). By local nonsatiation, for each n there is a yn satisfying
d(yn , y) < n1 and yn ≻ y. Since yn ≻ y and y ≽ x, transitivity implies yn ≻ x, so yn ∈ P (x). But
yn → y, so y ∈ P (x).

5 Proposition If ≽ is an upper semicontinuous and monotonic preference on Rn+ , then x ≧ y

implies x ≽ y.

Proof : Put xn = x + (1/n)1 and note that xn ≫ x ≧ y, so by monotonicity xn ≻ y, and thus

xn ∈ U (y). Now xn → x and by upper semicontinuity U (y) is closed, so x ≽ y.

Existence of utility
The following theorem is easy to prove, but the hypotheses are stronger than needed.

6 Theorem Let X = Rn+ and let ≽ be continuous and monotonic. Then ≽ has a continuous utility
function u.

Proof : Let 1 ∈ Rn be the vector whose components are all one, and let D = {λ1 : λ ⩾ 0} be the
“diagonal” of Rn+ . For x ∈ Rn+ , let M (x) = maxi xi and let m(x) = mini xi . Then M (x)1 ≧ x ≧ m(x)1,
so M (x)1 ≽ x ≽ m(x)1. Now let

L(x) = {λ ⩾ 0 : x ≽ λ1} H(x) = {λ ⩾ 0 : λ1 ≽ x}.

These sets are intervals (by monotonicity), closed (by continuity), and nonempty (since M (x)1 ∈ H(x)
and m(x)1 ∈ L(x)). Moreover since ≽ is total and reflexive H(x) ∪ L(x) = [0, ∞). Thus H(x) ∩ L(x)
consists of a single point λ(x)1 ∼ x. Define the function u : Rn+ → R+ by

u(x) is the unique number satisfying x ∼ u(x)1.

See Figure 1.
We now show that u is a utility for ≽. First assume that u(x) ⩾ u(y). Then u(x)1 ≧ u(y)1, so by
monotonicity u(x)1 ≽ u(y)1. But then x ∼ u(x)1 ≽ u(y)1 ∼ y, so by transitivity x ≽ y. For the converse,
assume x ≽ y. Then u(x)1 ∼ x ≽ y ∼ u(y)1, so u(x)1 ≽ u(y)1. Then by monotonicity, u(x) ⩾ u(y).
This completes the proof that u is a utility function for ≽.
We shall prove that u is continuous by proving that it is both upper and lower semicontinuous.3
For α ⩾ 0, {x ∈ Rn+ : u(x) ⩾ α} = U (α1), which is closed since ≽ is continuous. Similarly, the
set {x ∈ Rn+ : u(x) ⩽ α} = U −1 (α1), which is also closed. For α < 0, monotonicity implies that
{x ∈ Rn+ : u(x) ⩾ α} = Rn+ and {x ∈ Rn+ : u(x) ⩽ α} = ∅, again both closed. Thus u is both upper
and lower semicontinuous and therefore continuous.

Debreu [1, §4.6] proves that a continuous utility exists for a continuous preference on any connected
subset of Rn , without assuming monotonicity.
As a reminder, there are (total, reflexive, transitive) preferences on Rn+ that have no utility, but
which are discontinuous. Everyone’s favorite example is the lexicographic preference.4
3
There are many equivalent definitions of semicontinuity for functions. The easiest to use here is that u
is upper semicontinuous if for every α ∈ R, the set {x ∈ Rn+ : u(x) ⩾ α} is closed; and u is lower
semicontinuous if for every α ∈ R, the set {x ∈ Rn+ : u(x) ⩽ α} is closed. See section 17 of my on-line notes
on metric spaces for other characterizations.  
4
The lexicographic preference on the plane is given by (x, y) ≽ (x′ , y ′ ) if x > x′ or (x = x′ and y ⩾ y ′ ) . To

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 5

|
)1
(x
x

{z
H
M (x)1

}
u(x)1

}
m(x)1
L( {z
1
x)
|

Figure 1. Construction of utility for a monotonic continuous preference.

Convexity
When X is a subset of a linear space, we say that ≽ is

convex if y≻x =⇒ (1 − λ)x + λy ≻ x,

weakly convex if y≽x =⇒ (1 − λ)x + λy ≽ x,
strictly convex if y ̸= x & y ≽ x =⇒ (1 − λ)x + λy ≻ x,

for all 0 < λ < 1.

7 Lemma Let X be convex, and let ≽ be a preference on X. The following are equivalent.

1. ≽ is weakly convex.

2. For each x, the strict upper contour set P (x) is a convex set.

3. For each x, the weak upper contour set U (x) is a convex set.

Proof : (1) =⇒ (2) Assume ≽ is weakly convex, and let y1 and y2 belong to P (x). Since ≽ is
total and reflexive, without loss of generality we may assume that

y2 ≽ y1 ≻ x.
see that no utility exists for this preference relation, let x > x′ . Then any utility u would imply the existence of
rational numbers qx and qx′ satisfying





u (x, 1) > qx > u (x, 0) > u (x′ , 1) > qx′ > u (x′ , 0) .

This defines a one-to-one correspondence x ←→ qx between the reals and a subset of the rational numbers. But
Cantor proved long ago via his famous “diagonal procedure” that no such correspondence can exist.

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 6

Let 0 ⩽ λ ⩽ 1. By weak convexity then

(1 − λ)y1 + λy2 ≽ y1 ≻ x,

where the second relation follows from transitivity. Thus (1 − λ)y1 + λy2 belongs to P (x), which
proves that P (x) is convex.
(2) =⇒ (3) Assume that P (x) is convex for each x. Fix some x, and suppose by way of
contradiction, that U (x) is not convex. Then for some y1 and y2 belonging to U (x), that is,
y1 ≽ x and y2 ≽ x, and for some 0 < λ̄ < 1, the point z = (1−λ̄)y2 +λ̄y1 does not belong to U (x).
Then since ≽ is total, we must have x ≻ z. Thus y1 ≽ x ≻ z and y2 ≽ x ≻ z, so by transitivity
y1 ≻ z and y2 ≻ z. By hypothesis P (z) is convex, we must have (1 − λ̄)y2 + λ̄y1 = z ≻ z, a
contradiction. Therefore U (x) must be convex.
(3) =⇒ (1) Assume that U (x) is convex, and let y ≽ x. Then for any 0 < λ < 1, we have
(1 − λ)x + λy ∈ U (x), that is, (1 − λ)x + λy ≽ x, so ≽ is weakly convex.

Despite its name, the property of weak convexity is not actually weaker than convexity.

8 Exercise (Convexity does not imply weakly convexity) Let X = [−1, 1] and define
≽ by means of the utility function

1 x ̸= 0
u(x) =
0 x = 0.

Prove that ≽ is convex, but not weakly convex. □

The preference relation in the example above is not continuous, which brings up the next
lemma. To simplify the discussion of these properties let say that z is between x and y if
(i) x ̸= y, and (ii) z = (1 − λ)x + λy for some 0 < λ < 1.

9 Lemma (Convexity and upper semicontinuity imply weak convexity) If ≽ is con-

vex and upper semicontinuous, then it is weakly convex.

Proof : Assume that y ≽ x. In case y ≻ x, then by convexity (1 − λ)x + λy ≻ x for 0 < λ < 1,
so a fortiori (1 − λ)x + λy ≽ x.
So now consider the case y ∼ x and assume by way of contradiction that for some 0 < λ̄ < 1
the point z = (1 − λ̄)x + λ̄y satisfies x ≻ z. By upper semicontinuity, P −1 (x) is open, so we may
choose λ̃ close to λ̄, but with λ̃ > λ̄ so that w = (1 − λ̃)x + λ̃y satisfies x ≻ w. See Figure 2. But
this means that z is between w and x, and since x ≻ w, convexity implies z ≻ w. On the other
hand, w is between y and z, and y ∼ x ≻ z, so convexity implies w ≻ z, a contradiction.

10 Corollary Let X be convex, and let ≽ be a weakly convex regular preference on X. Any
utility function for ≽ is quasiconcave. If in addition, ≽ is convex, any utility function for ≽ is

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 7

z = (1 − λ̄)x + λ̄y
(1 − λ̃)x + λ̃y = w
x

Figure 2. (x ≻ w =⇒ z ≻ w) and (y ≻ z =⇒ w ≻ z). Oops.

explicitly quasiconcave.5
The next result gives conditions that rules out “thick” indifference classes.

11 Lemma If X is convex, and ≽ is convex, continuous, and nonsatiated, then P (x) is the
interior of U (x).

Proof : Since P (x) ⊂ U (x) and P (x) is open by lower semicontinuity, we have P (x) ⊂ int U (x).
For the reverse inclusion, let y belong to the interior of U (x), so there is some ε > 0 such that
the ε-ball centered at y lies wholly in U (x). Assume by way of contradiction that y ∈ / P (x).
Then since y ∈ U (x), it must be that y ∼ x. Since ≽ is nonsatiated, there is a point z ∈ X with
z ≻ y. Choose α < 0 but close enough to zero, so that the point w = (1 − α)y + αz is within
ε of y and also so that z ≻ w, which can be done by upper semicontinuity of ≽. See Figure 3.
Then z ≻ w ≽ x ∼ y. But since y lies between z and w, by convexity we must have y ≻ w, a
contradiction.

y w = (1 − α)y + αz

Figure 3. w ≽ x ∼ y and z ≻ y ≻ w, oops.

12 Example (Lemma 11 may fail without convexity) Let X = R and let ≽ be defined
by the utility u(x) = x2 . Then ≽ is locally nonsatiated and continuous, but P (0) = R \ {0} ̸=
R = int U (0). □
5
Recall that a function u : X → R is quasiconcave if for each α ∈ R, the set {x ∈ X : u(x) ⩾ α} is convex.
This is equivalent to the following condition: for all x, y in X and all 0 < λ < 1


f (1 − λ)x + λy ⩾ min{f (x), f (y)}.

A function u is explicitly quasiconcave if it is quasiconcave and in addition, for all x, y in X and all 0 < λ < 1


f (x) > f (y) =⇒ f (1 − λ)x + λy > min{f (x), f (y)} = f (y).

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 8

The next result is a riff on the one above, where convexity is replaced by weak convexity
and and nonsatiation is replaced by local nonsatiation.

13 Lemma If X is convex, and ≽ is weakly convex, lower semicontinuous, and locally nonsa-
tiated, then P (x) is the interior of U (x).

Proof : Since P (x) ⊂ U (x) and P (x) is open by lower semicontinuity, we have P (x) ⊂ int U (x).
For the reverse inclusion, let y belong to the interior of U (x), so there is some δ > 0 such
that the δ-ball V centered at y lies wholly in U (x). Assume by way of contradiction that
y∈/ P (x). Then since y ∈ U (x), it must be that y ∼ x. Since ≽ is locally nonsatiated, there is
a point z̄ ∈ X with z̄ ≻ y. Moreover by lower semicontinuity, there is some ε > 0 such that the
ε-ball B centered at z̄ is included in P (y).
Consider the open sets of the form

Aα = (1 − α)y + αB, where α < 0.

See Figure 4. The set Aα is a ball of radius |α|ε centered at w̄ = (1 − α)y + αz̄. Moreover,

z̄
B
y
w̄
Aα

Figure 4. The set A = (1 − α)y + αB, where α < 0.

∥w̄ − y∥ = |α|∥z̄ − y∥, so

w ∈ Aα =⇒ ∥w − y∥ ⩽ ∥w − w̄∥ + ∥w̄ − y∥ < |α| ε + ∥z̄ − y∥ .


So for |α| < δ/ ε + ∥z̄ − y∥ , we have ∥w − y∥ < δ, so Aα ⊂ V ⊂ U (x). So fix such an α < 0.
Now for any point w ∈ Aα there is by definition some z ∈ B such that
1 −α
w = (1 − α)y + αz, or y = w+ z.
1−α 1−α
I claim that this implies that w ∼ y. To see why, first note that since Aα ⊂ Bδ (y) ⊂ U (x) =
U (y) we have w ≽ y. But if w ≻ y, since z ≻ y, and y is a proper convex combination of z and
w, weak convexity implies y ≽ w ≻ y, a clear contradiction. This means that w ∼ y. But w is
an arbitrary element of the open set Aα , so local nonsatiation is violated on Aα .
This contradiction shows that int U (x) ⊂ P (x), completing the proof that int U (x) = P (x).

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 9

Lemma 9 gives conditions under which convexity implies weak convexity. The next one
gives conditions under which weak convexity implies convexity.

14 Lemma (Weak convexity, lower semicontinuity, and local nonsatiation imply convexity)
If X is convex, and ≽ is weakly convex, lower semicontinuous, and locally nonsatiated, then
it is convex.

Proof : Assume y ≻ x. Then y ∈ P (x) = int U (x) and x ∈ U (x), so by a well known property of
convex sets, the entire segment [y, x) lies in the interior of the convex set U (x). By Lemma 13,
int U (x) = P (x), so for 1 ⩾ λ > 0, we have (1 − λ)x + λy ≻ x. That is, ≽ is convex.

Preference maximization
15 Definition The point x∗ is a ≽-greatest point in the set B if x∗ ∈ B and for every x ∈ B,
we have x∗ ≽ x.

16 Lemma Let X be a finite nonempty set, and assume ≽ is a total, reflexive, and transitive
binary relation on X. Then X has a ≽-greatest element.

Proof : This provides an excellent opportunity to demonstrate proof by induction. Let P[n]
denote the proposition: If X has n elements, then X has a ≽-greatest element.
Clearly P[1] is valid, that is if X has a single element x, then by reflexivity, x ≽ x, so x is
≽-greatest in X.
We now show that P[n] implies P[n + 1]. So assume that X has n + 1 elements. Pick some
x ∈ X, and let A = X \ {x}. By the induction hypothesis P[n], since the set A has n elements,
it has a ≽-greatest element y. Since ≽ is total, there are two (overlapping) cases: x ≽ y and
y ≽ x.
Case 1: x ≽ y. Since y is ≽-greatest in A, we have y ≽ z for all z ∈ A = X \ {x}. By
transitivity we must have x ≽ z for z ∈ A, and by reflexivity x ≽ x. Thus x ≽ z for all z ∈ X
so x is ≽-greatest in X.
Case 2: y ≽ x. We already have y ≽ z for all z ∈ A = X \ {x}. Therefore y ≽ z for all
z ∈ X, so y is ≽-greatest in X.
We have just shown that P[n] implies P[n + 1]. By the principle of induction, every finite
set has a greatest element.

Alternative proof : Let |X| = m. Define y1 = x1 and inductively define


xn+1 if xn+1 ≽ yn ,
yn+1 =
 y if yn ≻ xn+1 .
n

Since preferences are total, each yn is well defined, and since preferences are transitive, the last
term ym satisfies ym ≽ xi for i = 1, . . . , m. (This actually requires a simple proof by induction.)
That is, ym is ≽-greatest in X.

17 Proposition Let X be nonempty and compact, and assume ≽ is upper semicontinuous

(and complete and transitive). Then X has a ≽-greatest element.

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 10

Proof : The family U = {U (x) : x ∈ X} consists of closed sets by upper semicontinuity. We

now show that U has the finite intersection property.
Let {x1 , x2 , . . . , xm } be a nonempty finite subset of X. By Lemma 16, {x1 , x2 , . . . , xm } has
T
a ≽-greatest element y. I claim that m i=1 U (xi ) = U (y). To see this note that y ≽ xi for each
i, so if z ∈ U (xi ), that is, if z ≽ y, then by transitivity z ≽ xi . That is U (y) ⊂ U (xi ) for each
T T
i, so U (y) ⊂ m U (xi ). On the other hand y = xi for some i, so U (y) ⊃ m i=1 U (xi ). Thus
Tm i=1
i=1 U (xi ) = U (y), as claimed. Since y ∈ U (y), this intersection is nonempty. In other words,
U has the finite intersection property.
Since X is compact, and U is a family of closed sets with the finite intersection property,
T
it has nonempty intersection. That is, there is some y belonging to x∈X U (x). But this just
says that y ≽ x for all x ∈ X.

18 Proposition Let X be convex, and assume that the regular preference ≽ is strictly convex.
Then a ≽-greatest element is unique (if it exists).

Proof : Suppose that both x and y are ≽-greatest elements of X. Suppose by way of contradic-
tion that if x ̸= y. Then by strict convexity, 21 y + 12 x ≻ x, contradicting the hypothesis that x
is ≽-greatest. Therefore x = y.

Preference maximization and expenditure minimization

Let X be a subset of Rn , and set β(p, w) = {x ∈ X : p · x ⩽ w}, where p ∈ Rn and w ∈ R.

19 Lemma Assume ≽ is locally nonsatiated and transitive.

Assume that
x∗ is ≽-greatest in β(p, w),
that is,
x∗ ∈ β(p, w) and p · x ⩽ w =⇒ x∗ ≽ x.
The contrapositive of the last statement is

x ≻ x∗ =⇒ p · x > w ⩾ p · x∗ .

Then

1. p · x∗ = w (all income is spent).

2. x∗ minimizes p · x over U (x∗ ). That is,

x ≽ x∗ =⇒ p · x ⩾ p · x∗ .

Proof : (1.) Assume by way of contradiction that p · x∗ < w. Then there is some ε > 0 such
that for all y with d(x∗ , y) < ε, we have p · y < w, so y ∈ β(p, w). By local nonsatiation, at
least one such one y satisfies y ≻ x, which contradicts the hypothesis that x∗ is ≽-greatest in
β(p, w).
Therefore p · x∗ = w.

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 11

(2.) Assume by way of contradiction that there is some x ∈ U (x∗ ) with p · x < p · x∗ . Then
there is some ε > 0 such that for all y with d(x, y) < ε, we have p · y < p · x∗ ⩽ w, so y ∈ β(p, w).
See Figure 5. By local nonsatiation, at least one such one y satisfies y ≻ x. But x ∈ U (x∗ )
means x ≽ x∗ , so by transitivity, y ≻ x∗ , which contradicts the hypothesis that x∗ is ≽-greatest
in β(p, w).
Therefore x∗ minimizes p · x over U (x∗ ).

y
ε x
x∗

Figure 5. Preference maximization subject to a budget implies expenditure minimization.

20 Lemma Assume X is convex, and ≽ is lower semicontinuous.

Assume x∗ minimizes p · x over U (x∗ ), and the cheaper-point assumption holds, that is,
there exists x̃ ∈ X satisfying p · x̃ < p · x∗ . Then x∗ is ≽-greatest in B = β(p, p · x∗ ).

Proof : Suppose by way of contradiction that there is some y ∈ B satisfying y ≻ x∗ , that is,
y ∈ P (x∗ ) ⊂ U (x∗ ). Then p · y ⩾ p · x∗ , as x∗ minimizes expenditure over U (x∗ ). But y is in
the budget B, so we conclude p · y = p · x∗ .

y
ε

x̃
y(λ) x∗

Figure 6. Expenditure minimization implies preference maximization

For λ satisfying 0 < λ < 1, define y(λ) = (1 − λ)y + λx̃. Then p · x̃ < p · y(λ) < p · x∗ . Since
X is convex, y(λ) ∈ B for all 0 < λ ⩽ 1.
But y(λ) → y as λ → 0, and y belongs to the open set P (x∗ ) (lower semicontinuity), so for
some ε > 0, for every λ < ε we have y(λ) ∈ P (x∗ ) ⊂ U (x∗ ). See Figure 6. But for such λ we
have p · y(λ) < p · x∗ , which contradicts the hypothesis that x∗ minimizes p · x over U (x∗ ).
Therefore x∗ is ≽-greatest in B = β(p, p · x∗ ).

To see what may happen if the cheaper-point assumption is violated, consider the following
example.

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 12

21 Example (Why the cheaper point is needed) Let X = R2+ . Let preferences be de-
fined by the utility function u(x1 , x2 ) = x1 +x2 . (This preference relation is continuous, convex,
and locally nonsatiated.) Let x∗ = (1, 0) and p = (0, 1). Then x∗ minimizes p · x over U (x∗ ).
But β(p, p · x∗ ) = β(p, 0), which is just the x1 -axis. This budget set is unbounded and no
≽-greatest element exists. See Figure 7.
If you don’t like the fact that I resorted to using a zero price, consider the case where X =
{x ∈ R2+ : x1 + x2 ⩾ 2}. Let u(x1 , x2 ) = 2x1 + x2 , p = (1, 1), and x̂ = (1, 1). Again x̂ minimizes
expenditure over U (x̂), but x̄ = (2, 0) is ≽-greatest in the budget set β(p, p · x̂) = β(p, 2), which
is the southwest boundary of the consumption set. See Figure 8. □

p
p X

β(p, 2)
U (x̂) x̂ U (x̂)
β(p, 0)
x̂ x̄
Figure 7. Cheaper-point viola- Figure 8. Cheaper-point viola-
tion 1. tion 2.

22 Corollary Assume X is convex, and ≽ is lower semicontinuous and locally nonsatiated.

Let p be given and set w = p · x∗ . Assume there is a point x̃ ∈ X satisfying p · x̃ < w.
Then x∗ is ≽-greatest in β(p, w) if and only if x∗ minimizes p · x over U (x∗ ).

Local nonsatiation, monotonicity, and demand

On the face of it local nonsatiation is a weaker condition than monotonicity for preference
relations. Clearly monotonicity implies local nonsatiation, provided the consumption set is
increasing. (A set A in Rn is increasing if (x ∈ A & z ≫ x) =⇒ z ∈ A.) But the preference
relation with utility u(x, y) = y − (x − 1)2 is locally nonsatiated without being monotone. See
Figure 9. But this preference relation generates the same demand as the preference relation
with utility given by 
y − (1 − x)2 x⩽1
û(x, y) = (1)
 y x⩾1
See Figure 9. I leave it to you to verify this claim.
This is true more generally. That is, if ≽ is a locally nonsatiated upper semicontinuous
preference, then there is a monotonic preference ≽∗ that generates the same demand.

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 13

Figure 9. Indifference curves for locally nonsatiated utility u(x, y) = y − (1 − x)2 and the
monotonic utility given by û(x, y) = y − (1 − x)2 for x ⩽ 1 and û(x, y) = y for x ⩾ 1.

To simplify the explicit description of such a monotonic relation, let us introduce the
following notation. Given a set A, write x ≽ A to mean “x ≽ y for all y ∈ A,” and let
D(x) = {y ∈ Rn+ : y ≦ x}.

23 Proposition Let ≽ be locally nonsatiated and upper semicontinuous regular preference on

Rn+ . Then the binary relation ≽∗ defined by

x ≽∗ y if there exists v such that x ≧ v ≽ D(y)

is a monotonic upper semicontinuous preference relation that generates the same demand for
(p, w) ∈ Rn++ × R+ . Moreover, if ≽ is convex, then ≽∗ is convex.

Proof : Note that for each y the set D(y) is compact, and since ≽ is upper semicontinuous, the
set µ(y) of ≽-greatest element of D(y) is nonempty. So x ≽∗ y if and only if there exist x ≦ x
and y ∈ µ(y) with x ≽ y.
First we show that ≽∗ is
i. reflexive: Let v ∈ µ(x). Then x ≧ v ≽ D(x), so x ≽∗ x.
ii. transitive: Assume x ≽∗ y and y ≽∗ z. Then there exist vy , vz , uy , and uz such that
x ≧ vy ≽ uy ∈ µ(y) and y ≧ vz ≽ uz ∈ µ(z).
Since vz ⩽ y and uy ≽ D(y) we have uy ≽ vz . Thus x ≧ vy ≽ uy ≽ vz ≽ D(z), so x ≽∗ z.
iii. total: If x ̸≽∗ y, then by definition, for every u ≦ x, there is some v ≦ y with v ≻ u.
Let u∗ ∈ µ(x), and let v ∗ ≦ y satisfy v ∗ ≻ u∗ . Thus y ≧ v ∗ ≻ u∗ ≽ D(x). Thus y ≽∗ x.

∗
We can now characterize the strict relation ≻ by
∗
y≻ x if and only if ∃v such that y ≧ v ≻ D(x).

Next we show that ≽∗ is monotonic, that is, if x ≫ y, then x ≻∗

y. So assume x ≫ y and let
z ∈ µ(y) ⊂ D(y), so x ≫ z. Then for ε > 0 small enough, d(v, z) < ε implies x ≫ v. By local
∗
nonsatiation, at least one such v satisfies v ≻ z ≽ D(y). Thus x ≫ z ≻ D(y), so x ≻ y.

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 14

To see that ≽∗ is upper semicontinuous, I shall prove that if y ≻∗

x, then there is an ε > 0
′ ∗ ′ ∗
such that d(x, x ) < ε implies y ≻ x too. So assume y ≻ x. Then there exists v ≦ y such

that
v ≻ D(x). Since D(x) is compact, I claim
there is some ε > 0 such that v ≻ Nε D(x) .6 Then
∗ ′
if d(x, x′ ) < ε, we have D(x′ ) ⊂ Nε D(x) too, so v ≻ D(x′ ). But this implies y ≻ x.

Finally we show that for p ≫ 0, a point x∗ is ≽-greatest in β(p, w) if and only if x∗ is also
∗
≽-greatest.
Assume first that x∗ is ≽-greatest in β(p, w). Let y ∈ β(p, w). Then D(y) ⊂ β(p, w), so
x ≧ x∗ ≽ D(y). Thus x∗ ≽∗ y. Therefore x∗ is ≽∗ -greatest in β(p, w).
∗

Now assume that x∗ is ≽∗ -greatest in β(p, w), and let x̄ be ≽-greatest. Since x∗ ≽∗ x̄ there is
some z ≦ x∗ with z ≽ D(x̄). In particular, z ≽ x̄ ∈ D(x̄), so z too is ≽-greatest. But by local
nonsatiation p · z = w, so z ≦ x∗ ∈ β(p, w) implies z = x∗ , so x∗ is also ≽-greatest.

To see that ≽∗ is convex if ≽ is convex, let x, x′ ≽∗ y, where x ≧ v ≽ D(y) and let x′ ≧ v ′ ≽

D(y). Then (1 − λ)x + λx′ ≧ (1 − λ)v + λv ′ and assuming ≽ is convex, (1 − λ)v + λv ′ ≽ D(y).
Thus (1 − λ)x + λx′ ≽∗ y.

Finally we mention that if u is an upper semicontinuous locally nonsatiated function, then

its monotonic hull v, defined by v(x) = max{u(y) : 0 ≦ y ≦ x}, is the smallest monotonic
function that dominates u. The Berge Maximum Theorem implies that it is upper semicontin-
uous. If u is a utility that represents ≽, then its monotonic hull represents ≽∗ .

24 Example Consider the quasilinear utility function for two goods x and y defined by

u(x, y) = y − (1 − x)2

which gives a linear demand function for x. It is locally nonsatiated but not monotone. It has
the property that the demand for x never exceeds 1. It has the same demand behavior as the
monotone utility 
y − (1 − x)2 x⩽1
v(x, y) =
 y x⩾1
which is its monotonic hull. See Figure 10. □

References
[1] G. Debreu. 1959. Theory of value: An axiomatic analysis of economic equilibrium. Num-
ber 17 in Cowles Foundation Monographs. New Haven: Yale University Press.
cowles.econ.yale.edu/P/cm/m17/m17-all.pdf
6

Here Nε (A) denotes the ε-neighborhood of A, that is, x : ( ∃y ∈ A ) [ d(x, y) < ε ] . It is an open set, being
the union of the open balls of radius ε centered on points of A. To see why this claim is true, let F denote the
closed upper contour set {u : u ≽ v}. Then the distance function d(z, F ) = inf{d(z, u) : u ∈ F } is (Lipschitz)
continuous and so achieves its minimum over the compact set D(x). Since F and D(x) are disjoint closed sets
this minimum is strictly greater than zero. Choose ε > 0 less than this minimum.

v. 2019.10.30::12.40
KC Border Technical Results on Regular Preferences and Demand 15

Figure 10. Indifference curves for locally nonsatiated utility u(x, y) = y −(1−x)2 and monotone
utility with same demand.

[2] A. Mas-Colell, M. D. Whinston, and J. R. Green. 1995. Microeconomic theory. Oxford:

Oxford University Press.

[3] M. K. Richter. 1971. Rational choice. In J. S. Chipman, L. Hurwicz, M. K. Richter, and H. F.

Sonnenschein, eds., Preferences, Utility, and Demand: A Minnesota Symposium, chapter 2,
pages 29–58. New York: Harcourt, Brace, Jovanovich.

v. 2019.10.30::12.40