CALIFORNIA INSTITUTE OF TECHNOLOGY

Division of the Humanities and Social Sciences

Introduction to Correspondences
KC Border
September 2010
Revised November 2013
v. 2013.11.25::15.25

In these particular notes, all spaces are metric spaces. The set of extended real numbers,
R {, }, is denoted R .

Correspondences

There are many instances when a set of points depends on a parameter. For instance:
In economics, the budget set (p, w) = {x Rn+ : p x w} is the set of shopping
lists x that can be purchased with wealth w at the price list p. This set depends on the
parameter vector (p, w) Rn+ R+ .
In a metric space (X, d), the open ball B (x) = {y X : d(y, x) < } is a set that depends
on the parameter list (x, ) X R++ .
For a family of constrained optimization problems
maximize f (x) subject to g(x) = ,
the constraint set {x X : g(x) = } depends on the parameter .
In order to capture this dependence we would like to have a notion of a set-valued function.
The seemingly obvious idea of a function f : X 2Y from the set X into the power set (the set
of subsets) of Y may not be the best choice. The problem comes when we try to imagine its
graph, which is a subset of X 2Y . A simpler idea is what we call a correspondence, which is
just another name for a binary relation from X to Y .
Definition 1 A correspondence from X to Y associates to each point in X a subset (x)
of Y . We write this as : X Y . For a correspondence : X Y , let gr denote the graph
of , which we define to be
gr = {(x, y) X Y : y (x)}.
Let : X Y , and let F X. The image (F ) of F under is defined to be

(F ) =
(x).
xF

The value (x) is allowed to be the empty set, but we call {x X : (x) = }, the domain
of , denoted dom .
The terms multifunction, point-to-set mapping, and set-valued function are also used
for a correspondence.
Depending on your spatial relation skills, it may be possible to visualize the graph of a
correspondence as we have defined it, especially if X and Y are subsets of the line or plane.

KC Border

Introduction to Correspondences

Remark 2 Technically a singleton-valued correspondence : X Y is not a function from X

to Y , since the value of is a subset of Y with only element, e.g., {y}, whereas the value of a
function is point y in Y . But the graph of a singleton-valued correspondence is also the graph
of a function f , where (x) = {f (x)}, and we may find it convenient to identify and f (that
is, treat them as though they are the same object). We may also use an expression like the
correspondence is a function to mean that it is singleton-valued.

Inverse images

For correspondences there are two natural notions of inverse.

Definition 3 The upper (or strong) inverse of E under , denoted u [E], is defined by
u [E] = {x X : (x) E}.
The lower (or weak) inverse of E under , denoted [E], is defined by
[E] = {x X : (x) E = }.
For a single y in Y , define

1 (y) = {x X : y (x)}.

Note that 1 (y) = [{y}].

Berge [8] uses the notation + and , and Hildenbrand [11] uses and 1 for the upper
and lower inverses respectively.

Notions of continuity for functions

Let (X, dX ) and (Y, dY ) be metric spaces, and let f : X Y be a function. When it is clear
to which space a pair of points belongs, I may drop the subscripts on dX and dY and simply
write d for the
Remember that Rn is a metric space under the Euclidean (or 2 ) metric
metric.
n
n
2
d2 (x, y) =
i=1 (xi yi ) . Other popular metrics on R include
nthe max-metric (or )
d (x, y) = maxi |xi yi | and the taxicab metric (or 1 ) d1 (x, y) = i=1 |xi yi |. For a review
of the important concepts (such as open sets, neighborhoods, closed sets, compact sets, and
(semi)continuous functions) relating to metric spaces see my on-line notes or Hildenbrand [11,
Chapter 1]. The two standard, equivalent, definitions of continuity are these.
Definition 4 (Continuity using neighborhoods) A function f : X Y is continuous
at x if for every neighborhood G of f (x), its inverse image f 1 [G] is a neighborhood of x. That
is, for every open set G with f (x) G, there exists an open set U with x U such that
z U = f (z) G.
Note that this definition is topological, that is, it uses only the notion of neighborhoods and
does not mention the metrics. The next definition makes explicit use of the metrics.
Definition 5 (Continuity using metrics) A function f : X Y is continuous at x if
for every > 0 there exists some > 0 such that
(
)
dX (x, z) < = dY f (x), f (z) < .

v. 2013.11.25::15.25

KC Border

Introduction to Correspondences

The following proposition is a simple consequence of the definition of an open set in a metric
space, but lets walk through it to see where problems arise when trying to define continuity for
correspondences.
Proposition 6 A function f is continuous at x in the topological sense of Definition 4 if and
only if it is continuous at x in the metric sense of Definition 5.
Proof : ( = ) Assume f( is continuous
at
)
( x in the
) topological sense (Definition 4), and let > 0
be given. Then G = B f (x) = {y : d y, f (x) < } is an open set with f (x) G, so there is
some open set U containing x such that z U = f (z) G. By definition of openness of U ,
there is a > 0 such that dX (z, x) < = z U . Thus
(
)
dX (z, x) < = dY f (z), f (x) < ,
that is, f is continuous at x in the metric sense (Definition 5).
( = ) Assume f is continuous at x in the metric sense (Definition 5), and let G be
( an open
)
set containing f (x). By definition of openness of G, there is some >(0 such that) dY y, f (x) <
= y G. Then there is a > 0 such that dX (x, z) < = dY f (z), f (x) < . That is,
letting U denote the open set B (x), we have
z U = f (z) G,
so f is continuous at x in the topological sense (Definition 4).

Notions of continuity for correspondences

To apply Definition 4 or Definition 5 to correspondences we first have to confront the problem

that there are two natural notions of inverse image, so there are two natural versions of each
definition. But there is another problem, namely, the topological and metric versions are not
equivalent. But first we have to figure out the appropriate statement of the metric - definition.
Define the distance function d(x, A), where A is a set and x is a point, by
d(x, A) = inf d(x, z).
zA

Then d is continuous. In fact it satisfies the Lipschitz condition



d(x, A) d(z, A) d(x, z).
Definition 7 Let A be a set in a metric space (X, d). The -neighborhood N (A) of A is
defined by1
N (A) = xA B (x) = {x X : d(x, A) < }.
Note that as the union of the open balls B (x), the set N (A) is open. If X is also a vector
space with translation-invariant metric2 d, then we also have N (A) = A + B (0).
Now lets start with the upper inverse. The - definition of upper hemicontinuity is this:
1 Borwein

and Zhu [10, p. 3] call this set, with strict inequality replaced by weak inequality, the -enlargement
of A. It is possible to put useful topologies on 2X , in which case the term neighborhood might be confusing. But
I dont intend to topologize 2X here.
2 A metric d on a vector space is translation-invariant if for all x, y, z, we have d(x, y) = d(x + z, y + z).

v. 2013.11.25::15.25

KC Border

Introduction to Correspondences

Definition 8 (Metric upper hemicontinuity)

A correspondence : X Y satisfies
Property UM at x if the upper inverse of an -neighborhood of (x) is a neighborhood of
x. In other words, satisfies Property UM at x if for every > 0, there is > 0 such that
(
)
d(z, x) < = (z) N (x) .
We say that satisfies Property UM if it satisfies Property UM at each x X.
The topological version is this:
Definition 9 (Upper hemicontinuity) A correspondence : X Y satisfies Property UT
at x if the upper inverse of an open neighborhood of (x) is a neighborhood of x. In other words,
satisfies Property UT at x if for every open set G with (x) G, there is an open set U with
x U such that
z U = (z) G.
We say that satisfies Property UT if it satisfies Property UT at each x X.
What is the relationship between these properties?
Proposition 10 If : X Y satisfies Property UT at x, then it satisfies Property UM at x.
If : X Y satisfies Property UM at x and if (x) is compact, then it satisfies Property UT
at x.
The first half of the proposition is straightforward and similar to Proposition 6, the second
half relies on the following lemma.
Lemma 11 Let G be an open subset of a metric space (X, d) and let K be a nonempty compact
subset of G. Then there is an > 0 such that
K N (K) G.
Proof : If K G, where G is open, then Gc is closed and disjoint from K. The distance function
x 7 d(x, Gc ) is continuous, so it achieves a minimum on the compact set K at some point x
K.
But we must have d(
x, Gc ) > 0, because if d(
x, Gc ) = 0, then since Gc is closed, we would have
x
Gc K contradicting the fact that K and Gc are disjoint. Any with 0 < < d(
x, Gc )
satisfies the conclusion.
Thus Property UT is stronger than Property UM for correspondences that do not have
compact values, and the next example shows that it may be so strong as to rule out many
natural correspondences.
Example 12 (Cf. Aubin and Ekeland [3, p. 108] for a related example.) The correspondence
m
: Rm
+ R defined by
(x) = {y Rm : y x}
satisfies Property UM , but does not satisfy Property UT .
To see that this violates Property UT , let n = 2 and let
G = {(x, y) R2 : y < 1/|x| if x < 0}.
(
)
Note that( this )set includes the entire right half-plane as well as (0, 0) . But for any > 0,
the set (, ) contains (x, /2), which does not belong to G when x < 0 and |x| > 2/. Thus
violates Property UT .

v. 2013.11.25::15.25

KC Border

Introduction to Correspondences

Ill leave it to you to show that does satisfy Property UM . Hint: If x x < and y x,
then y = y + x x satisfies y y < and y = x (x y) x .
This example shows that it might be impossible for a correspondence satisfy Property UT
if is not compact-valued. Observe that a constant correspondence (that is, (x) = A for all x)
does satisfy Property UT , even if A is not compact.

So either Property UT or UM can be used as the basis for a definition of continuity. In the
past Ive preferred UT (e.g., [1, 9]), but now I may wish to join other authors, e.g., Aubin [2]
or Aubin and Ekeland [3, p. 108], who prefer UM . Nevertheless, until I get around around to
rewriting everything, I will stick with the following definition.
Definition 13 A correspondence : X Y is upper hemicontinuous (uhc) at x if it
satisfies Property UT at x; and is lower hemicontinuous (lhc) at x if whenever x is in
the lower inverse of an open set so is a neighborhood of x.
In other words, is uhc at x if for any open set G Y , if
(x) G,
then there is an open set U X containing x such that
z U = (z) G.
And is lhc at x if for any open set G Y , if
(x) G = ,
then there is an open set U X containing x such that
z U = (z) G = .
The correspondence : X Y is upper hemicontinuous if it is upper hemicontinuous
at every x X. The correspondence : X Y is lower hemicontinuous if it is lower
hemicontinuous at every x X. Thus is upper hemicontinuous if the upper inverses of open
sets are open and is lower hemicontinuous if the lower inverses of open sets are open.
A correspondence is continuous if it is both upper and lower hemicontinuous.
The term semicontinuity is used by some authors to mean hemicontinuity (indeed Aubin [2]
or Aubin and Ekeland [3], Phelps [14], and Borwein and Zhu [10] use it). The term hemicontinuiuty has been prevalent in mathematical economics since the appearance of Hildenbrand [11].
Warning! The definition of upper hemicontinuity is not fully agreed upon. For example,
Berge [7, 8] requires in addition that have non-empty compact values in order to be called
upper hemicontinuous. Hildenbrand [11, Definition 1, p. 21] requires nonempty values. See
Moore [13] for a catalog of related definitions. It is true that the most interesting results apply
to compact-valued correspondences (e.g., compare Examples 12 and 24 below), but it seems
useful to me and others, including Beer [6, Definition 6.2.4, p. 193], Borwein and Zhu [10,
Definition 5.1.15, p. 173], Hildenbrand [11, Definition 1, p. 21], and Phelps [14, Definition 7.2,
p. 102], to make upper hemicontinuity and compact values separate properties. These authors
use the term usco as an adjective to describe a correspondence that is upper hemicontinuous
and compact-valued.
If : X Y is singleton-valued, the upper and lower inverses of a set coincide and agree with
the inverse regarded as a function. Either form of hemicontinuity is equivalent to continuity as
a function. Thus a semicontinuous real-valued function is not a hemicontinuous correspondence
unless it is also a continuous function. It is also possible to put topologies on the space of
compact subsets of Y that capture hemicontinuity by regarding as a function into 2Y . See
Klein and Thompson [12, Theorems 7.1.4 and 7.1.7, pp. 7375].
v. 2013.11.25::15.25

KC Border

Introduction to Correspondences

Definition 14 The correspondence : E F is closed at x if whenever xn x, yn (xn ),

and yn y, then y (x). A correspondence is closed if it is closed at every point of its
domain, that is, if its graph is closed.
Fact 15 If is closed at x, then (x) is a closed set.
Proof : Let yn be a sequence in (x) with yn y. Define xn = x for all n, so xn x and
yn (xn ). Since is closed at x, we have y (x). This shows that (x) is a closed set.
Example 16 (Closedness vs. upper hemicontinuity) In general, a correspondence may
be closed without being upper hemicontinuous, and vice versa.
Define : R R via
{ 1
x = 0
{x}
(x) =
{0}
x = 0.
Then is closed but not upper hemicontinuous.
Define : R R via (x) = (0, 1). Then is upper hemicontinuous in my sense, but not
closed.3

However, a closed-valued upper hemicontinuous correspondence is a closed correspondence,

and if the range space Y is compact, then a closed correspondence is upper hemicontinuous.
Proposition 17 Let : X Y be a upper hemicontinuous at x and assume that (x) is a
closed set. Then is closed at x.
Moreover, if Y is compact, and is closed at x, then is upper hemicontinuous at x.
Proof : Let be upper hemicontinuous at x and assume that (x) is closed. Let xn x,
yn (xn ), and yn y. Suppose by way of contradiction that y
/ (x). Since (x) is
closed, there are disjoint open sets G and U in Y with (x) G and y U . Since is upper
hemicontinuous at x, the upper inverse u [G] is an open neighborhood of x. Since xn x, there
is some N such that n N implies xn u [G] or (xn ) G. In particular, yn (xn ) G,
so yn
/ U . Thus yn y, a contradiction. This establishes that is closed at x.
Now assume that Y is compact and is closed at x. Assume by way of contradiction that
there is some open set G with (x) G but u [G] is not a neighborhood of x. This means
that for every n there is some xn in B1/n (x), the ball of radius 1/n centered at x, such that
(xn ) G. Thus there exists yn (xn ) with yn
/ G. Since Y is compact, the sequence yn has
a subsequence converging to some y Y . Along this subsequence we have xn x, yn (xn ),
yn y so since is closed at x we must have y (x) G. But no yn G, which contradicts
yn y. This contradiction establishes that for every open set G with (x) G it must be the
case that u [G] is a neighborhood of x. That is, is upper hemicontinuous at x.
Remark 18 We can weaken the assumption that the range space Y is compact. What we need
is for each x there is a neighborhood whose image is included in some compact set. So let us
say that is locally bounded at x if there is a neighborhood U of x in X and a compact set
K Y such that (U ) K. Then the argument above proves the following result.
Corollary 19 Let : X Y be locally bounded and closed at x. Then is upper hemicontinuous at x.
Corollary 20 Let : X Y .
3 Again,

under Berges definition, an upper hemicontinuous correspondence is automatically closed.

v. 2013.11.25::15.25

Draw pictures.

KC Border

Introduction to Correspondences

1. If is upper hemicontinuous and closed-valued, then is closed.

2. If is closed and locally bounded everywhere, then is upper hemicontinuous.
Proposition 21 (Hemicontinuity and sequences) Let : X Y .
1. If has nonempty compact values, then is upper hemicontinuous at x if and only if for
every sequence xn x and yn (xn ) there is a convergent subsequence of yn with limit
in (x).
2. Then is lower hemicontinuous at x if and only if xn x and y (x) imply that there
is a sequence yn (xn ) with yn y.
Proof : (1.) ( = ) Suppose is upper hemicontinuous at x, xn x and yn (xn ). Since
is compact-valued, (x) has a bounded neighborhood U . Since is upper hemicontinuous,
there is a neighborhood V of x such that (V ) U . Thus yn is eventually in U , thus bounded,
and so has a convergent subsequence. Since compact sets are closed, this limit belongs to (x).
( = ) Now suppose that for every sequence xn x, yn (xn ), there is a subsequence
of yn with limit in (x). Suppose is not upper hemicontinuous; then there is a neighborhood
U of x and a sequence zn x with yn (zn ) and yn
/ U . Such a sequence yn can have no
subsequence with limit in (x), a contradiction.
(2.) ( = ) Assume is lower hemicontinuous at x and that xn x and y (x).
Then for each natural number k, the set Gk = {z Y : d(z, y) < 1/k} is an open set with
y Gk (x) = . Since is lower hemicontinuous at x, the lower inverse [Gk ] is a
neighborhood of x. Since xn x, there is some n(k) such that for all n n(k), we have
xn [Gk ], that is, (xn ) Gk = . Without loss of generality we may take n(k + 1) > n(k)
for all k. For n < n(1) pick any yn (xn ). For n(k) n < n(k + 1) pick yn (xn ) Gk .
Then n n(k) implies d(yn , y) < 1/k, so yn y as desired.
( = ) Assume that whenever xn x and y (x), there is a sequence yn (xn ) with
yn y. Let G be an open subset of Y with (
x) G = , and let y (
x) G. Let U = [G].
We need to show that U is a neighborhood of x
. Suppose by way of contradiction that U is not
a neighborhood of x
. Then for each natural number n, the ball B1/n (
x) is not a subset of U , so
pick xn B1/n (
x) U c . Then xn x
, but (xn ) G = for each n. On the other hand, by
hypothesis there are yn (xn ) with yn y. Thus there is some n0 such that n n0 implies
yn G. But then for n n0 , we have yn n G = , a contradiction. Therefore U = [G]
is a neighborhood of x
.
Proposition 22 Let , : X Y , and define ( ) : X Y pointwise by ( )(x) =
(x) (x). Suppose (x) (x) = .
1. If and are upper hemicontinuous at x and closed-valued, then the correspondence
( ) is upper hemicontinuous at x.
2. If is closed at x and is upper hemicontinuous at x and (x) is compact, then ( )
is upper hemicontinuous at x.
Proof : Let U be an open neighborhood of (x) (x). Put C = (x) U c .
(1) Note that C is closed and (x) C = . Thus there are disjoint open sets V1 and V2
with (x) V1 and C V2 . Since is upper hemicontinuous at x, there is a neighborhood W1
of x with (W1 ) V1 V2c . Now (x) U V2 , which is open and so x has a neighborhood
W2 with (W2 ) U V2 , as is upper hemicontinuous at x. Put W = W1 W2 . Then for
z W , (z) (z) V2c (U V2 ) U . Thus ( ) is upper hemicontinuous at x.
/ (x)
(2) Note that in this case C is compact and (x) C = . Since is closed at x, if y
then we cannot have yn y, where yn (xn ) and xn x. Thus there is a neighborhood Uy
v. 2013.11.25::15.25

KC Border

Introduction to Correspondences

of y and Wy of x with (Wy ) Uyc . Since C is compact, we can write C V2 = Uy1 Uyn ;
so setting W1 = Wy1 Wyn , we have (W1 ) V2c . The rest of the proof is as in part (1).

Exercises and examples

Example 23 Every constant correspondence, (x) = A for all x, is both upper and lower
hemicontinuous.

m
m
m
Example 24 Let Rm
+ = {x R : x 0}. The compact-valued correspondence : R+ R+
defined by
(x) = {y Rm
+ : 0 y x}

is continuous. (Compare this with Example 12.)

For this argument it is most convenient to use the max metric on Rm . (But the Euclidean
metric also works.) We start by observing the following fact.

Fact 25 Let x, x Rm
+ satisfy d(x, x ) < (recall that d is the max metric),
0 y x and set
y = (y + x x)+

(where for any vector z Rm , the vector z + is defined to be the coordinate-wise

maximum of z and 0). See Figure 1. Then
d(y, y ) <

and

y x .

B (y)

B (x)

y
y + x x

0
Figure 1. 0 y x and d(y, y ) < .

Proof : Consider the ith coordinate yi . There are two cases:

Case (i): yi > 0. Then yi = yi + xi xi , so |yi yi | = |xi xi | < and
yi = xi + (yi xi ) xi .
Case (ii): yi = 0 and yi + xi xi 0. The latter inequality only occurs if xi xi ,
in which case we have > xi xi yi = |yi yi |. Also, yi = 0 xi .
So in either case |yi yi | < and yi xi .
u
To see that is upper hemicontinuous, let G be an open subset of Rm
+ and assume x [G].

We need to find a > 0 so that if d(x, x ) < , then (x ) G.

Just as in the proof of Lemma 11, the distance function d(z, Gc ) achieves a minimum value
m on the compact set (x). Since (x) and Gc are disjoint closed sets the minimum value m

v. 2013.11.25::15.25

KC Border

Introduction to Correspondences

(
)
satisfies m (> 0. )Set = m/2. Then the neighborhood N (x) of (x) is disjoint from Gc ,
that is, N (x) G.
By the above fact (interchanging x and x ), for x B (x), if z x , then z lies within
of some point z x (namely z = (z + x x )+ ), so z G. Thus (x ) G.

To prove the lower hemicontinuity of , let G be an open subset of Rm

+ and assume x [G].

That is, there is some y G with 0 y x. We need to find a > 0 so that if d(x, x ) < ,
then (x ) G = .
m
Since G is open relative to Rm
+ , there is some > 0 such that B (y) R+ G, where

B (y) = {z Rm
+ : max |yi zi | < }.
i

By the fact above there is some y G (x ). This proves the lower hemicontinuity of at
x.

Exercise 26 If the graph of is open, then is lower hemicontinuous.

Exercise 27 Let f, g : X R be continuous and assume that f (x) g(x) for all x. Show that
x 7 { R : f (x) g(x)}

is continuous.
Example 28 The classical budget space is
B = {(p, w) Rn R : p 0, w > 0}.
Define B0 = {(p, w) Rn R : p 0, w 0}. The budget correspondence
(p, w) = {x Rn+ : p x w}

is compact-valued and continuous on B. It is closed, but not lower hemicontinuous on B0 .

Upper hemicontinuity on B follows from the fact that the graph of is clearly closed, and
it is easy to see that is locally bounded on B.
For lower hemicontinuity at a point (
p, w)
B, assume G is open and let x
(
p, w)
G.
There are two cases: Case 1 is that x
= 0. In this case, pick some x
= x
satisfying x
x
and
x
G. Then p x
< w,
since p 0. Thus U = {(p, w) B :
px
< w} is an open neighborhood
of (
p, w)
such that for every (p, w) U , we have x
(p, w) G = , which shows that is lhc
at (
p, w).
Case
2
is
that
x

=
0.
In
this
case,
U
=
B is an open neighborhood of (
p, w)
satisfying

x
(p, w) G = , for all (p, w) U . This proves that is lhc at (
p, w).

Now in case w
= 0 and some pj = 0, then x
= ej (the j th unit coordinate vector) we have
x
(
p, 0). Now consider the sequence (pn , wn ) (
p, 0) defined by pn = p + (1/n)ej and
wn = 1/n2 . If x (pn , wn ), then xj 1/n, so if xn (pn , wn ) we cannot have xn x
, so
by Proposition 21, is not lhc at (
p, 0).

The Berge maximum theorem

One of the most useful and powerful theorems employed in mathematical economics and game
theory is the maximum theorem due to Berge [7]. It deals with the continuity of the solution
and value of a parametrized family of constrained optimization problems. We start with a set
P of parameter or state values, and a set X of controls or actions or choice variables. To each
parameter value p, there is a constraint set or feasible set (p) of controls,
: P X.
v. 2013.11.25::15.25

KC Border

Introduction to Correspondences

10

The goal is to maximize an objective function f over the feasible set. We allow f to depend on
both the parameter and the control. The domain of f need not be all of P X, but it must
include the graph of ,
f : gr R.
We define the optimal value function V : P R by
V (p) = sup{f (p, x) : x (p)}.
Note that we allow this to be extended real-valued (the supremum may be ). And recall that
sup = .
Definition 29 An extended real-valued function f : Z R on a metric space is upper semicontinuous if for every R, the upper contour set {z Z : f (z) } is closed (or equivalently {z Z : f (z) < } is open). The function f is lower semicontinuous if for every
R, the lower contour set {z Z : f (z) } is closed (or equivalently {z Z : f (z) > }
is open).
We can localize the concept and say that f is upper semicontinuous at z if f (z) < implies
that there is a neighborhood U of z such that x U implies f (x) < . Lower semicontinuity at
a point is defined similarly.
You should prove the following.
Fact 30 A real-valued function is continuous if and only if it is both upper and lower semicontinuous.
Proposition 31 (Lower semicontinuity of the optimal value function) Let : P X
be lower hemicontinuous at p0 . Let f : gr R be lower semicontinuous. Then the optimal
value function V is lower semicontinuous at p0 .
Proof : We need to prove that if V (p0 ) > , then there is a neighborhood W of p0 such that
p W implies V (p) > . So assume that V (p0 ) > . Then there is some x0 (p0 ) satisfying
f (p0 , x0 ) > .
Since f is lower semicontinuous on gr , there is an open neighborhood U V of (p0 , x0 ) such
that for all (p, x) gr U V we have f (p, x) > . Since is lower hemicontinuous,
W = [V ] U is a neighborhood of p0 . By the definition of lower inverse, for each p W
there is some x (p) V , which implies f (p, x) > . Therefore V (p) = supx(p) f (p, x) >
too.
Proposition 32 (Upper semicontinuity of the optimal value function)
Assume that
the feasibility correspondence : P X has nonempty compact values. Let f : gr R be
upper semicontinuous. Then the optimal value function V is actually a maximum,
V (p) = max{f (p, x) : x (p)}.
If is upper hemicontinuous at p0 , then V is upper semicontinuous at p0 .
Proof : We need to show that if V (p0 ) < , then there is a neighborhood W of p0 such that
for all p W we have V (p) < . Since V (p0 ) < , for every x (p0 ) we have f (p0 , x) < .
Since f is upper semicontinuous, for each x (p0 ), there is a neighborhood Wx Ux of (p0 , x)
such that for every (p, z) Wx Ux we have f (p, z) < . Since (p) is compact, there are

v. 2013.11.25::15.25

KC Border

Introduction to Correspondences

finitely many x1 , . . . , xk such that U =

j=1

Uxj includes (p0 ). Put W =

11
k

j=1

Wxj . Then for

any (p, x) W U we have f (p, x) < . Moreover (p0 ) U , so u [U ] is a neighborhood of

p0 . Now for any p W u [U ] and any x (p) we have x U so f (p, x) < which implies
V (p) < . That is, V is upper semicontinuous at p0 .
Berge Maximum Theorem
Let : P X be a compact-valued correspondence. Let
f : gr R be continuous. Define the argmax correspondence : P X by
(p) = {x (p) : x maximizes f (p, ) on (p)},
and the optimal value function V : P R by
V (p) = f (p, x)

for any x (p).

If is continuous at p, then is closed and upper hemicontinuous at p and V is continuous at

p. Furthermore, is compact-valued.
Proof : Given the previous Propositions 31 and 32, all that remains is to show that is closed
at p, for then = and Proposition 22(2) implies that is upper hemicontinuous at p.
Let pn p, xn (pn ), xn x. We wish to show x (p). Since is upper hemicontinuous
and compact-valued, Proposition 21(1) implies that indeed x (p). Suppose x
/ (p). Then
there is z (p) with f (p, z) > f (p, x). Since is lower hemicontinuous at p, by Proposition 21
there is a sequence zn z with zn (pn ). Since zn z, xn x, and f (p, z) > f (p, x), the
continuity of f implies that eventually f (pn , zn ) > f (pn , xn ), contradicting xn (pn ).

More about correspondences

Proposition 33 (Upper hemicontinuous image of a compact set) Let K be a compact

set and assume that : K Y is upper hemicontinuous and compact-valued. Then (K) is
compact.
Proof : Let {U } be an open covering of (K). Since (x) is compact, there is a finite subcover
Ux1 , . . . , Uxnx , of (x). Put Vx = Ux1 Uxnx . Then since is upper hemicontinuous, u [Vx ]
is open and contains x. Hence K is covered by a finite number of u [Vx ]s and the corresponding
Uxi s are a finite cover of (K).
Proposition 34 (Sums of correspondences) Let i : X Rm , i = 1, . . . , k.
(a) If each i is upper hemicontinuous at x and compact-valued, then

i : z 7
i (z)
i

is upper hemicontinuous at x and compact-valued.

(b) If each i is lower hemicontinuous at x, then the sum

i is lower hemicontinuous at x.

Proof : Exercise. Hint: Use Proposition 21(1).

Proposition 35 (Products of correspondences) Let i : E Yi , i = 1, . . . , k.

v. 2013.11.25::15.25

KC Border

Introduction to Correspondences

12

(a) If each i is upper hemicontinuous at x and compact-valued, then

i : z 7
i (z)
i

is upper hemicontinuous at x and compact-valued.

(b) If each i is lower hemicontinuous at x, then the product
x.

(c) If each i is closed at x, then i i is closed at x.

i is lower hemicontinuous at

Proof : Exercise.

References
[1] C. D. Aliprantis and K. C. Border. 2006. Infinite dimensional analysis: A hitchhikers guide,
3d. ed. Berlin: SpringerVerlag.
[2] J.-P. Aubin. 1998. Optima and equilibria: An introduction to nonlinear analysis, 2d.
ed. Number 140 in Graduate Texts in Mathematics. Berlin, Heidelberg, & New York:
SpringerVerlag.
[3] J.-P. Aubin and I. Ekeland. 1984. Applied nonlinear analysis. Pure and Applied Mathematics: A Wiley-Interscience Series of Texts, Monographs, and Tracts. Mineola, New York:
John Wiley & Sons. Reprint of the 1984 edition by John Wiley and Sons.
[4] J.-P. Aubin and H. Frankowska. 1990. Set-valued analysis. Boston: Birkhuser.
[5] G. Beer and P. Kenderov. 1988. On the arg min multifunction for lower semicontinuous
functions. Proceedings of the American Mathematical Society 102(1):107113.
http://www.jstor.org/stable/2046040
[6] G. A. Beer. 1993. Topologies on closed and closed convex sets. Number 268 in Mathematics
and Its Applications. Dordrecht: Kluwer Academic Publishers.
[7] C. Berge. 1959. Espaces topologiques et fonctions multivoques. Paris: Dunod.
[8]

. 1997. Topological spaces. Mineola, New York: Dover. Originally published in

French by Dunod, Paris, 1962 as Espaces topologiques, fonctions multivoques. Reprint
of the English translation by E. M. Patterson, originally published by Oliver and Boyd,
Edinburgh and London, 1963.

[9] K. C. Border. 1985. Fixed point theorems with applications to economics and game theory.
New York: Cambridge University Press.
[10] J. M. Borwein and Q. J. Zhu. 2005. Techniques of variational analysis. Number 20 in
CMS Books in Mathematics/Ouvrages de Mathmatiques de la SMC. New York: Springer
Science+Business Media.
[11] W. Hildenbrand. 1974. Core and equilibria of a large economy. Princeton: Princeton
University Press.
[12] E. Klein and A. C. Thompson. 1984. Theory of correspondences: Including applications to
mathematical economics. New York: John Wiley and Sons.

v. 2013.11.25::15.25

KC Border

Introduction to Correspondences

13

[13] J. C. Moore. 1968. A note on point-set mappings. In J. P. Quirk and A. M. Zarley,

eds., Papers in Quantitative Economics, 1, pages 129137. Lawrence, Kansas: University
of Kansas Press.
[14] R. R. Phelps. 1993. Convex functions, monotone operators and dierentiability, 2d. ed.
Number 1364 in Lecture Notes in Mathematics. Berlin: SpringerVerlag.

v. 2013.11.25::15.25