Escolar Documentos
Profissional Documentos
Cultura Documentos
In this section we review some elements of convex analysis to be used in the proof of
Nashs theorem. The detailed proofs are given for the convenience of the readers. The
readers are also referred to the books [2, 6] for more complete study of convex analysis in
finite dimensions. Throughout the note we consider the Euclidean space Rn with the inner
product denoted by h, i and the Euclidean norm denoted by k k.
For two points a and b in Rn , the line segment connecting them is
[a, b] := a + (1 )b [0, 1] .
Note that if a = b, then this interval reduces to a singleton [a, b] = {a}.
A subset of Rn is convex if [a, b] whenever a, b . Equivalently, is convex if and
only if a + (1 )b for all a, b and [0, 1].
Let f : Rn R be an real-valued function. The epigraph of f is a subset of Rn R defined
by
epi f := (x, ) Rn+1 x Rn and f (x) .
The function f is called convex if
f x + (1 )y f (x) + (1 )f (y) for all x, y Rn and (0, 1).
From the definition, we can show that f is convex if and only if its epigraph is a convex set.
Given a nonempty set Rn , the distance function associated with is defined by
d(x; ) := inf kx k , x Rn .
(1.1)
For each x Rn , the Euclidean projection from x to is defined based on the distance
function as follows
P(x; ) := kx k = d(x; ) .
(1.2)
1
Fariborz Maseeh Department of Mathematics and Statistics, Portland State University, Portland, OR
97202, United States (mau.nam.nguyen@pdx.edu). This note is based on a presentation at the Analysis
Seminar at Portland State. The research of Nguyen Mau Nam was partially supported by the Simons
Foundation under grant #208785.
Proposition 1.1 Let be a nonempty closed subset of Rn . Then for any x Rn , the
Euclidean projection P(x; ) is nonempty.
Proof. By definition (1.2), for each k N, there exists k such that
1
d(x; ) kx k k < d(x; ) + .
k
Obviously, {k } is a bounded sequence. Thus it has a subsequence {km } that converges
to . Since is closed, . Letting m in the inequality
d(x; ) kx km k < d(x; ) +
1
,
km
< kx 1 k2 = d(x; ) ,
= kx 1 k2
2
4
1 + 2
.
2
Next we characterized the Euclidean projection to convex sets in Rn .
which is a contradiction due to the inclusion
P(x1 ; ) P(x2 ; )
2 P(x1 ; ) P(x2 ; ), x1 x2 .
In particular, it implies the Lipschitz continuity of the projection mapping with constant
` = 1:
P(x1 ; ) P(x2 ; )
kx2 x2 k for all x1 , x2 Rn .
(1.4)
Proof. It follows from the preceding proposition that
P(x2 ; ) P(x1 ; ), x1 P(x1 ; ) 0 for all x1 , x2 Rn .
Changing the role of x1 , x2 in the above inequality and summing them up give us
P(x1 ; ) P(x2 ; ), x2 x1 + P(x1 ; ) P(x2 ; ) 0.
3
This implies the first estimate in the proposition. Finally, the nonexpansive property of the
Euclidean projection follows directly from
P(x1 ; ) P(x2 ; )
2 P(x1 ; ) P(x2 ; ), x1 x2
kP(x1 ; ) P(x2 ; )k kx1 x2 k
for all x1 , x2 Rn , which completes the proof of the proposition.
(1.5)
The following proposition establish a useful relation between the normal cone and the
projection to convex sets.
Proposition 1.5 Let be a nonempty closed convex subset of Rn and let x
. Then
v N (
x; ) if and only if x
P(
x + v; ).
Proof. By the definition, one has that v N (
x; ) if and only if
hv, w x
i = h
x+vx
, w x
i 0 for all w ,
which is equivalent to the fact that x
P(
x + v; ) by Proposition 1.3.
This implies
f (
x + t(u x
)) f (
x)
0.
t
Taking the limit as t 0+ , one obtains
hf (
x), u x
i 0,
which can be rewritten as hf (
x), u x
i 0. Since this inequality holds true for any
u , one has that f (
x) N (
x; ).
Let us prove the converse. Suppose that f (
x) N (
x; ). Then
0 hf (
x), u x
i for all u .
Since f is a convex function, we always have
hf (
x), u x
i f (u) f (
x) for all u Rn .
In particular, for every u ,
0 hf (
x), u x
i f (u) f (
x).
Thus, f (
x) f (u) for every u , which ensures that x
is an optimal solution of (P).
Nash Equilibrium
This section gives a brief introduction to noncooperative game theory and presents a simple
proof of the existence of Nash equilibrium as a consequence of convexity and Brouwers
fixed point theorem. For simplicity, we only consider two-person games while more general
situations can be treated similarly.
Definition 2.1 Let 1 and 2 be nonempty subsets of Rm and Rn , respectively. A noncooperative game in the case of two players I and II consists of two strategy sets i
and two real-valued functions ui : 1 2 R for i = 1, 2 called the payoff functions.
Then we refer to this game as {i , ui } for i = 1, 2.
The key notion of Nash equilibrium was introduced by John Forbes Nash, Jr. in 1950 who
proved the existence of such an equilibrium and was awarded by the Nobel Memorial Prize
in Economics in 1994; see, e.g., [8] and the references therein for more details.
Definition 2.2 Given a noncooperative two-person game {i , ui }, i = 1, 2, a Nash equilibrium is an element (
x1 , x
2 ) 1 2 satisfying the conditions
u1 (x1 , x
2 ) u1 (
x1 , x
2 ) for all x1 1 ,
u2 (
x1 , x2 ) u2 (
x1 , x
2 ) for all x2 2 .
5
Player I
Player II
A
4, 4
1, 3
3, 1
2, 2
In this example, Nash equilibrium occurs when both players choose strategy A, and it
also occurs when both players chooses strategy B. Let us consider, for instance, the case
where both players choose strategy B. In this case, given that Player II chooses strategy
B, Player I also wants to keep strategy B because a change of strategy would lead to a
reduction of his/her payoff from 2 to 1. Similarly, given that Player I chooses strategy
B, Player II wants to keep strategy B because a change of strategy would also lead to a
reduction of his/her payoff from 2 to 1.
Example 2.4 Let us consider another simple two-person game called matching pennies.
Suppose that Player I and Player II each has a penny. Each player must secretly turn the
penny to heads or tails and then reveal his/her choices simultaneously. Player I wins the
game and gets Player IIs penny if both coins show the same face (heads-heads or tails-tails).
In the other case where the coins show different faces (heads-tails or tails-heads), Player II
wins the game and gets Player Is penny. The payoff function of each player is represented
in the payoff matrix below.
Player I
Player II
Heads
Tails
Heads
1, 1
1, 1
Tails
1, 1
1, 1
In this game it is not hard to see that no Nash equilibrium exists. Indeed, denote by A
the matrix that represents the payoff function of Player I, and denote by B the matrix that
6
1 1
1 1
and B =
1 1
1 1
Now suppose that Player I is playing with a mind reader who knows Player Is choice of
faces. If Player I decides to turn heads, then Player II knows about it and chooses turn
tails to win the game. In the case where Player I decides to turn tails, then Player II
chooses to turn heads and again wins the game. Thus to have a fair game, he/she decides
to randomize his/her strategy by, for instance, tossing the coin instead of putting the coin
down. We describe the new game as follows.
Player I
Player II
Heads, q1 Tails, q2
Heads, p1
1, 1
1, 1
Tails, p2
1, 1
1, 1
In this new game, Player I uses a coin randomly with probability of coming up heads p1
and probability of coming tails p2 , where p1 + p2 = 1. Similarly, Player II uses another
coin randomly with probability of coming up heads q1 and probability of coming tails q2 ,
where q1 + q2 = 1. The new strategies now are called mixed strategies while the original
ones are called pure strategies. Now suppose that Player II uses mixed strategy {q1 , q2 },
then Player Is expected payoff for playing the pure strategy heads is
uH (q1 , q2 ) = q1 q2 = 2q1 1.
Similarly, Player Is expected payoff for playing the pure strategy tails is
uT (q1 , q2 ) = q1 + q2 = 2q1 + 1.
Thus Player Is expected payoff for playing mixed strategy {p1 , p2 } is
u1 (p, q) = p1 uH (q1 , q2 ) + p2 uT (q1 , q2 ) = p1 (q1 q2 ) + p2 (q1 + q2 ) = pT Aq,
where p = [p1 , p2 ]T and q = [q1 , q2 ]T .
By the same arguments, if Player I chooses mixed strategy {p1 , p2 }, then Player IIs expected
payoff for playing mixed strategy {q1 , q2 } is
u2 (p, q) = pT Bq.
For this new game, an element (
p, q) is a Nash equilibrium if
u1 (p, q) u1 (
p, q) for all p ,
u2 (
p, q) u2 (
p, q) for all q ,
where := [p1 , p2 ]T p1 0, p2 0, p1 + p2 = 1 is a nonempty, compact subset of R2 .
7
Nash [3, 4] proved the existence of his equilibrium in the class of mixed strategies in the
setting of Example 2.4 and in more general setting. His proof was based on Brouwers
fixed point theorem as well as other, rather involved arguments. Now we present a much
simpler proof of the Nash equilibrium theorem (due to Rockafellar) that also uses Brouwers
fixed point theorem while applying in addition just elementary tools of convex analysis and
optimization; see [7].
Theorem 2.5 Consider a two-person game {i , ui }, where 1 Rm and 2 Rm are
nonempty compact convex sets. Let the payoff functions ui : 1 2 R be given by
u1 (p, q) := pT Aq,
u2 (p, q) := pT Bq,
where A and B are an m n matrices. Then this game admits a Nash equilibrium in the
sense of Definition 2.2.
Proof. It follows from the definition that an element (
p, q) 1 2 is a Nash equilibrium
of the game under consideration if and only if
u1 (p, q) u1 (
p, q) for all p 1 ,
u2 (
p, q) u2 (
p, q) for all q 2 .
It follows from Proposition 1.5 that this holds if and only if we have the normal cone
inclusions
p u1 (
p, q) N (
p; 1 ), q u2 (
p, q) N (
q ; 2 ).
(2.6)
By using the structures of ui , conditions (2.6) can be equivalently expressed as
A
q N (
p; 1 ) and B T p N (
q ; 2 ),
or (A
q , B T p) N (
p; 1 )N (
q ; 2 ) = N ((
p, q); 1 2 ) = N ((
p, q); ), where := 1 2 .
By Proposition 1.3 on the Euclidean projection, this is equivalent to
(
p, q) = P((
p, q) + (A
q , B T p); ).
(2.7)
References
[1] A. Jofre, R. T. Rockafellar and R. J-B Wets, A variational inequality scheme for determining an economic equilibrium of classical or extended type, in Variational Analysis
and Applications (F. Giannessi and A. Maugeri, eds.), Springer, 553578, 2005.
8