NPS-53-85-0001

NAVAL POSTGRADUATE SGHOOl

Monterey, California

Pari- Mutue 1 as a Syst em of

Aggre gat ion of Information

by

Guillermo Owen
Oct ober 1984

Technical Report for Period

March 1984 - September 1984

Approved for public release; distribution unlimited

Prepared for: Chief of Naval Research

Arlington, VA 22217
PedDocs
D 208.14/2
NPS-53-85-0001
 NAVAL POSTGRADUATE SCHOOL
Monterey, California 93943

R. SHUMAKER
H. D. A. SCHRADY
.

Commodore, U.S. Navy Provost

Superintendent

Reproduction of all or part of this report is authorized

This report was prepared by:

 Unclassified DUDLEY KNOX LIBRARY
SECURITY CLASSIFICATION OF THIS PACE (When Data Entered) MAVAI PrtftTfiPAHi lATg g-CHOdi
READ INSTRUCT; 9394 3-51 Q-<
REPORT DOCUMENTATION PAGE BEFORE COMPLETING
1. REPORT NUMBER 2. GOVT ACCESSION NO. 3. RECIPIENT'S CATALOG NUMBER

NPS-53-85-0001
5. TYPE OF REPORT & PERIOD COVERED
4. TITLE (and Subtllte)
Pari-Mutuel as a System of Technical Report
Aggregation of Information March 1-Sept. 30, 1984
6. PERFORMING ORG. REPORT NUMBER

7. AUTHORS 8. CONTRACT OR GRANT NUMBERC»J

Guillermo Owen
PERFORMING ORGANIZATION NAME AND ADDRESS 10. PROGRAM ELEMENT. PROJECT. TASK
9.
AREA 4 WOHK UNIT NUMBERS
Naval Postgraduate School 6]]52N;RR000-0]-]0
Monterey, CA 93943 N000]484WR4]00]
II. CONTROLLING OFFICE NAME AND ADORESS 12. REPORT DATE

NPS Foundation Research Program October 1984

NUMBER OF PAGES
NPS, Monterey, CA 93943
13.

23
AGENCY NAME ADDRESSfff different from Controlling Ottice) 15. SECURITY CLASS. (o( thla report)
U. MONITORING ft

Chief of Naval Research

Arlington, VA 22217 Unclassified
15a DECLASSIFICATION/ DOWNGRADING
SCHEDULE

16. DISTRIBUTION STATEMENT (of thla Report)

Approved for public release; distribution unlimited

17. DISTRIBUTION STATEMENT (of the abatract entered In Block 20. If different from Report)

16. SUPPLEMENTARY NOTES

19. KEY WORDS (Continue on ravaraa aide If neceaeary and Identify by block number)

Group Decision
Group Probability
Pari-Mutuel
20. ABSTRACT (Continue on reveree aide If neceaeary and Identity by block number)
optimal
A bettor is faced with the problem of choosing
outcomes of an
bets given his (subjective) probabilities of Conversely,
experiment and the payoff odds on these outcomes
a bookie (pari-mutuel system), faced
by several bettors with
of choosing
different subjective probabilities, has the problem shown
It is
payoff odds so as to avoid the risk of loss.
under certain reasonably broad conditions on the several
that

DD ,^r73 1473 EDITION OF 1 NOV 65 IS OBSOLETE Unclassified

SECURITY CLASSIFICATION OF THIS PAGE (When Date
S.'N 0102- LF-014-6601 Entered)
 Unclassified
SECURITY CLASSIFICATION OF THIS PAGE (Wht\ Datm Entered)

bettors' utilities and subjective probabilities, such an

equilibrium set of payoff odds always exists. Some
examples are worked out in detail.

S N 0102- LF-014-6601
Unclassified
SECURITY CLASSIFICATION OF THIS PAGE(TW»»n Dmtm Enffd)
 DUDLEY KNO;
JWjVAL POSTG^:.
MONTEREY CA 93943-5101
Pari-Mutuel as a System of Aggregation

of Information

1. The Individual Bettor

Let us consider the problem faced by a bettor at a race

track. There are n horses, with (subjective) probabilities

P , p ,...,p of winning. He has A dollars to bet on these

horses, and a bet of x dollars on horse j will return x/q .

dollars if that horse wins. (The x/q . includes the bettor's

original x dollar bet.) It is assumed that the p. and a

3 3

satisfy the standard conditions

n
(1)
p .>

3=1
and

(2) £ q. = l ; g i
3=1

Condition (1) means the bettor's subjective probabilities

are consistent; condition (2) means the payoff odds are fair.

(Though, in fact, they seldom are, and most bookies will normally

announce odds such that \ q . is substantially greater than

1, i.e., they pay less than a "fair" system would.)

The bettor has a utility function u for money; it will be

assumed that u is monotone non-decreasing and continuous. The

bettor's problem, then, is to choose his bets, x ,...,x n , so as to

maximize his expected utility, given as

n

(3) f(x
i v 2 p . U (
-3- - B + A \ .
where
n

(4) b =
^ x .

3=1

subject to

(5) b < A

(6) x . > .

The first thing we notice is that, assuming the fairness

conditions (2), the bettor might as well set B = A, i.e. bet all

his available funds. In fact, suppose we had B < a. We could

then set e = A - B , and

x '. = x . + e q . j = 1 , . . . , n
3 3 3

In this case,

B'=>x.'=B
Iv + e

and, moreover,

F(x , ,

x _. + eq
F(x '
, . .
L r

so that f(x') = F(x), i.e. the bettor can do at least as well with

bets x .' such that b' = a as with any other bets with b < a.

We simplify the problem, then, to one of maximizing

(? ) ^(x x ) = yp u (1±
y
.
1 n 3 \q.
 .

subject to

(8)
J,-
(9) x . > .

Assuming differentiability of u, the first-order conditions

for optimality will be

(10a) -2- u> [-2.]= x if x . > o

Pi
(10b) -^- w (0) < \ if x . =

where X is a Lagrange multiplier representing the marginal utility

of money.

In case u is not dif f erent iable at the point x./q., conditions

J J

(10) must be modified, in terms of the right-hand and left-hand

derivatives of u to give

Now (10a) can be rewritten as

In the simplest case, all x. are positive, so that (10a) holds for

all j. Adding with respect to j , we have, by (2)

(12) X - £
j
P, *' ^
j

so that X is simply the expected value of u '

More generally, of course, (12) does not hold for all j, and
 . ^

so we can only state that X is at least equal to the expected

value of u '

In case u is concave, the first-order conditions (10) are

sufficient for optimality. We rewrite these as

u'
/ x
—M=
. \

—
\q .

if x . >

(14) u' (0) < ^_± if x .

=
P
J

Using the fact that u' is monotone non-increasing (for concave

u) , we obtain the interesting result

(15)
x
i
-^-
q
>
x
—
qk
k
whenever
p
—i
q
<*- > —
pk
qk
j j

with the stronger result that, for strictly concave u, (15) holds

even if the second inequality is loose.

Thus, a discrepancy between the bettor's subjective probabilitie

and the payoff odds leads the bettor to bet so that his conditional

winnings will be greater for horses for which the ratio P-/q.
is greater, and conversely.

Conditions (13) and (14) are meaningful if both p. and q.

are positive. In case p = 0, q. > 0, it is easily seen that

optimality requires x. = 0, i.e. never bet on a horse which

(subjectively) has no chance of winning. It is not clear what

happens if g. = o, though in practice it is difficult to imagine

a situation in which infinite odds were offered. In case p. =

J

q. = 0, we imagine the bettor will still set x. = , in case

P. > q. = 0, however, we seem to reach some sort of contradiction.

We note, then that q. = o leads to contradictions which would best

 ) i

be avoided; among other things, the payoff functions are discontinuous

or fail to exist here.

In case u is strictly concave, we may use the inverse function

w m (W)' 1 and (13) - (14) now take the form

(16) „. . 9j „(^i) if „(ifl) > o

(17) x . = if w I —
J

Condition (8) can be restated as

Xg
(is) , - ^ q . » (^i
J

where the prime on the summation svmbol means that it should con-

sider only those j such that (16) holds, i.e. such that

(19) \q . < p . W (0
The right side (18) can be seen to be a monotone non-increasing

function of \ and thus (18) can be solved, numerically or

analytically, for A. This presumably solves the single bettor's

problem.
2. The Equilibrium Odds

In general , bookies tend to be risk-averse and seek to set

payoff odds in such a way as to eliminate the possibility of loss.

Of course, a bookie is not bound by the fairness condition (2),

so that, in practice, the sum of the q. is greater than 1. If

(2) were to be enforced, however (perhaps under cutthroat competi-

tion among bookies), the bookie could only eliminate the risk of

loss if the amounts bet on the several horses were proportional

to the q . , i.e. if

(20) b . = q . c
J J

where b . is the total amount bet on horse j (by all bettors) and

c is the total amount of all wagers.

If there is only one bettor, it is easy to see that this can

he accomplished by setting q. = p.. For then x. = q. a will

J J J J

satisfy conditions (10) (with A = u'(A)). In case u is strictly

concave, moreover, this is the bettor's unique optimum, so that

q. = p. gives rise to an equilibrium. (Clearly, with one bettor,

b . = x . and c = A ) .

J J

If there are two or more bettors, the bookie must look for

some way of combining the several bettors' subjective probabilities

so as to avoid risk. At a race track, this is normally accomplished

by a pari-mutuel system, which simply sets q. = b./c, so that (20)

is automatically achieved, after the amounts bet are known. In

effect, the players bet against each other, with the track as

intermediary. This has the disadvantage - from the players' point

of view - that bets are made with only partial knowledge of the
 .

payoff odds. Thus, a player might well feel he would have changed

his bets, had he known the true payoff odds in advance. Of course

such a change would in turn cause the q. to change, leading to a

further change in bets, et sic ad infinitum ,

or at least until

some equilibrium is reached. The question is whether such an

equilibrium exists.

Assume, then, m bettors. Bettor i (i = 1,..., m) has a

subjective probability distribution (p.., p. 2 ,..., p. ) satisfying

p . .
> , and

I >v - ' .

j=l

This same bettor has a sum of money, a., available for betting,

and a utility function for money, u . . If the odds are posted as

(<7
7
' <?-,/•••/? ), then each bettor will choose (x , x ,..., x. )

so as to maximize his expected utility, as discussed above. Total

bets on horse j are then

m
(2D b . = y x .

i=l

and the total amount bet on all horses is

m n

(22) C = 1 *i' 1 b .

J
i=l J=J

There will be an equilibrium if (20) holds for all j.

As was mentioned above, difficulties arise if q . = for

J

any j. We will therefore try to avoid this, and will specifically

rule out such equilibria. We make then the following assumption.

Assumption z . For every j, there is some i such that p. . >

 We prove the existence of equilibrium under the further

assumption that the utility functions are strictly concave.

Essentially, this uses a fixed-point theorem. Some care must

however be used to avoid the possibility that the fixed point

lies on the boundary of the simplex.

Theorem 1 . Suppose Assumption z holds, and suppose more-

over that all the utility functions are strictly concave. Then

there is an equilibrium n-tuple of payoff odds, q .* > 0.

Proof : Let Q be the unit /i-simplex, i.e. the set of

vectors (g ,..., q ) satisfying (2). Let Q be the interior of

Q (the set of q with all components positive) and let 9£> be the

boundary of Q (the set of q with at least one q. = 0.)

*
o
For q c Q , consider bettor i's optimal choice of bets.

As discussed above, it cannot be optimal for him to bet on a

horse with no chance of winning, so his bets must satisfy, not

just (8) and (9), but also the condition x., = o whenever p., =
ik ik
Restricted to that set, bettor i's expected utility,
n
x
\.
i
(x.
i
f q) = )
t-d
p.
i
.

7
U .

i
(
\
—
a
J

is strictly concave, and so has a unique maximizing vector,

x
i
*(q). Since f is continuous for all x. and all q e Q° , it

will follow that x.*(q) is continuous for q e Q° .

Let , now,
m
b*(q) = y x.*(q)
i=l
Then b* is a continuous mapping from q° into n
iR . Let, finally,
 .

b
k *(q)
Z> .* (q)
J(q) = < j = ma x

Clearly, J assigns to each q € Q a non-empty subset of

n = \l, 2,..., n\ . By the continuity of b* , J is upper semi-

continuous .

Next, for q e dQ , define

J(q) = | J q = o\ .

Since q e dQ, J(q) is non-empty here also. Trivially, it is upper

semi-continuous if restricted to dQ

In this way, the mapping j is defined over the entire simplex

Q. We wish to show it is upper semi-continuous, i.e. if q —»§

and j e J(q), then j e J(q).

In this, we can dispense with the case in which q e Q , since

such q can only be approached through q e Q , and we know j, res-

tricted to I? , is semi-continuous. Similarly, we can dispense

with the case that q — -q, with all q and q e dQ, since we know

J, restricted to dQ , is semi-continuous.

It remains to consider the case in which q -g with q e Q

and q e dQ . Let k = J(q). We must show that, for q sufficiently

Close to q , J(q) <= K

Take some (fixed) k e K : we have q = . By assumption z,

K.

there is some bettor, b, with p > o. Keeping h fixed, let L (q)

be the set of all j for which p, ./a. is maximal.

Suppose q. > 0. As q g, the ratio p,,/q.

nk x.
increases without
j

bound, whereas p ./q . approaches the finite limit, p, /a . Thus

n J J ftj J

j \ l(q), and we conclude there exists c > such that, if

 Let r be the minimum of all a . such that q . > . Let
J 3

z = r/2 . Then, for all g such that \q - g| < z and all j i K,

we will have q. > r/2.

J

Let s be the minimum of all p, . such that a . = and p L . >

Since u is strictly increasing and concave, we know u '

(y + ) >

for all y. Set, then,

Finally, let e'4

4nC

Let now e be the smallest of z , z , z , and z . Assume

3

q e Q , \q - q\ < z. We will show J(q) c K.

Let x * (q) be bettor h's optimal response to the payoff odds

q, and suppose x .* > q./2 for some j k K. Let £ e l(q). Since

I? - $1 < e
2
/ i e K. Moreover, j i L(q), so
p h$ > p hj
q q
l j

and hence, by (15),

*
X
ll >

* *
Thus x £ and x are both positive, and so we can apply (13) to
h

set
*

V ('M\ -iLlhl v (
x
-m
Now, since lg - g| < e, we will have q. > r/2, q. < z and

10
 * 1
Also, p.. < 1. Finally, x,./g. > — , so by monotonicity

v (
x
Ai\ *
v <j>- Thus

and, using the definition of e

By the monotonicity of u ' , this gives

*
X
hi > 2C
Q* r

7T 7T 7T

Clearly, £>,i > X * , and so bq /q * > 2

*
On the other hand, for any k 4 K, b < c and q > r/2. Thus

and we see that k i J(g) . We conclude that J(q) c k.

*
Suppose, on the other hand, there is no j, j 4 k, with x. > q./2

In this case,

j i X j t K

and so

. * 1
Thus there is some I a K with x, > =—
hin 2n .

Since |
q - q |
< e^ and q. = 0, we have q. < E and so

X
'Z ^ h/ > i = 2C

11
 Once again b.*/q. < 2c/r for all j 4 K, and so b*/q. < b»*/q.

Thus we conclude once that J(q) c k.

We see then that J is an upper semi-continuous mapping,

assigning a non-empty subset of N to each q c Q. Define, now,

for s c n,
q e Q
$(S) = q
I
q . = if j £ S > .

J
)

Clearly, $ is an upper semi-continuous mapping from the sub-

setsof n to Q. The composition, * = <j>oJ, is then an upper semi-

continuous mapping assigning a non-empty, closed convex subset of

Q to each q e q . By the Kakutani fixed-point theorem, such a

mapping must have a fixed point, i.e. there is q e Q, such that

q e $ (q ) .

Clearly q* k dQ since, for q c dQ , J (q) consists of those

indices j with q . = 0, and so §(q) will consist of those vectors

z e Q such that z . = whenever q . > 0. Thus q* e 0° . But, if

q* r q , the only s c n such that q* c $ ( s) is n itself, i.e.

b .* (q*
—1—
)

J(q*) = n. This moans that is equal for all j, and

q
j
this will mean that
b .* (q* ) = q* C
J J

Thus q* is the desired equilibruim odds vector.

The hypothesis of Theorem 1 - mainly, strict concavity - is

overly restrictive. We can weaken it to require only (weak) con-
cavity together with strict monotonicity of the utility functions

Assume, then, that the u are merely concave functions of

±

money. For t > 0, define

w =
-X
. ( x , t ) u . (x ) - t

Then w (x,0) = u.(x). We have, however,

i

12
 "

X
w .
'
(x , t) = U .
'
(x) + t e

X
w .
(x ,t) = U ." (x) - t e

(where, in all cases, the primes denote differentiation with

respect to x) and so we find w. is strictly concave and monotone

in x for each t > .

Suppose, then, that each bettor's utility function u.(x) is

replaced by the strictly concave w.(x;t). For each t > 0, there

is an equilibrium n-tuple q*(t). As t -0, these g*(t) will have

an accumulation point, q**. The only difficulty is that, though

all q*(t) c Q ,
q** could conceivably belong to dQ. Theorem 2

says this will not happen.

Theorem 2. Assume all the functions u. are concave, strictly

i

monotone increasing, and suppose Assumption Z holds. Then there

exists an equilibrium n-tuple q * * e q .

Proof : As discussed above, let q*(t) be the equilibrium

obtained with the utility functions w . ( x , t ) . Let t ,- by the

compactness of Q, the points q*(t) will have some accumulation

point q** c Q. We must show a** i DC-

Suppose, then, q e dQ . We will show that, if |

q - <? |
and t are

sufficiently small, q * q*(t).

In fact, if q = q*(t), then J (a) = n, where J (q) is the

set J, as described above, with u. replaced bv w.(x,t).

i i

Define e , e , c , e , as in the proof of Theorem 1, and let

e = e /2 . Let S < U .
'
(j).
5

Suppose, now, that e is the minimum of z , e , e , e , that

\q - q\ < e, and that < t < 6. We have, n ow

13
 .

2C

v (*M -v (*f) *• r
>v (*)

v (H - v (i)
*<* 4 <v (i) «<«* (i)

and so

£
4
!i ££ • v (ff) » v (fM
The proof then proceeds as in that of Theorem 1, leading to the

conclusion that
J (?) c K * n
t

and so q * q* (t) . Thus q * q**, i.e. q** e Q° as desired.

-It is easy to see that, if q*(t) is an equilibrium for each

t, then q** will be an equilibrium for t = 0, i.e. for the utility

functions u .

It is possible to weaken the conditions on the utility functions

further - so that they are not strictly increasing, but only if we

strengthen condition z. Consider then

Assumption i
y. For all i and j, p. •>
. > 0.
i j

Theorem 3 . Assume all the utility functions are concave and

non-decreasing, and suppose Assumption y holds. Then there is an

equilibrium q* * e Q .

Proof : We consider first the trivial case in which, for each

bettor i, u.(x) is maximized at x = A. (or less). In this case,

i i '

it is clear that, for any q e q° , x.. = a .q . will be optimal for

bettor i, and so every q e q° will give an equilibrium.

14
 .

Suppose, then, that for some bettor, say bettor 1, u, is not

maximized at ^,. Then u' (A ,+)> 0, and we will set

U[ (Aj +)
r =
U (A ~ }
'l 1

Clearly < r < 1 .

As before, we replace the utility functions u.(x) by

w.(x,t) = u.(x) - t e '

, and consider the equilibria q* (t) obtained
in this manner. (Theorem 1 guarantees their existence.) As t '0,

these q*(t) have an accumulation point q** c Q. If q * * e Q , it is

the desired equilibrium. We must show q* * k 3£>

In fact, assume q e dQ , and let k be the set of all k such

that q = 0. Let e > be such that, if \q - q~\ < e,

k

^ ij < r ik

j k

for all i, all k e K, and j i K. (This can be done since all

By (15), and since r < 1, we have

X X
-^
. . . .

(24) ~ -J*
qk q
j

for all i, all j i K, k e K.

Since (24) holds for all i, the equilibrium condition (20)

can hold only if (24) holds as an equation throughout. This will

mean x_
±j = A j = 1 , . . . , n

Then, for k c K, j i k,

P lj < r P ±k
q 2 qk
j .

15
 ,

and, by (11)
V
w '
(A

(A
i
+!
-;
t}

t)
< r
2
i 2

or
Al
U ' (A. + ) + t e
1 I < r
U ' (A}-) + t e~ 1

so that, since r < 1,

V ' (A 1+ )
<
U '
(Al-) 2
2

and this is a contradiction. Thus q cannot be q*(t) for any t,

and therefore q * q**. We conclude that q* * e Q , and is the

desired equilibrium.

3. Examples

We consider here several examples. The first three consider

some "reasonable" utility functions; the last shows that the con-

ditions of Theorems 2 and 3 cannot be further weakened.

(a) Logarithmic utility

Assume that each bettor has a utility function

U . (x) = log (K . + x)
l l

where k . is a parameter, representing perhaps player i's reserves

In this case, the optimability conditions (10) - (11) take the

form
ij =A.^ if x . . >
K .q . + x . .
1J
i J 13

or, equivalently
P
25 13
( ) *
i7 - = ~y~ - K.q if positive

(26)
X = if p . . < K q X
i.7 ij .

*
.

3
.

1G
 .

In the simplest case, when all x . . > 0, this will give us

n

A
1
2
- z
*—>

3=1
*.
XJ
- -i
A
. — .
*.
2

and so
1
X .

2 A . + K .

2 2

Substituting in (25), this gives us

x . . = (A . + K . ) p 27. . - K .q .

1J 1 2 2 J

To look for the equilibrium, we have, from (20) - (21)

m
> x . . = q .

i=l

Thus
C
Zj (A
i
.

'
+ K
"l
.

"^
)p . - q
^
* Zj K x
7
. = q .*
J

so _
(27) (A + K )p ij
_ j j _ i
*j* = C + £K.
2

But c = lL a , and we see that q .

* is then simply a weighted
i J

average, with weights A. + K. f of the several subjective pro-

babilities p . . .

(b) Exponential utilities

Consider, next, the exponential utility functions

-a x .

U (x) = -e i

where a . is a parameter, measuring, in some sense, bettor i's

risk aversion. (Essentially, 1/a . can be thought of as representing

the sum of money which i would be "hurt" by losing.)

17
 In this case, the optimality conditions (10)-(11) take the

form
- — \ = X .q . if x . . >

which reduces to
P ij
(28) a . x . . = q . (log a . - log X . ) + q . log
i xj j l i l q

if this (right
° side) is positive, and x . .
= otherwise
2J

(i.e. x . . = if a. p.. < X .q . ) .

ij l ij i j

Assuming, again, that all x. . > 0, summation with respect

to j gives us
p
a
11
. A . = log a.
i
- log X . + y
/. q" . lo< n
7 J

so that, substituting in (28),

P
^
,-'
Ik i

q.
k
=
a
1
.

i
1 og
qk
- 1 q. log + A
j

For the equilibrium odds, we add with respect to i,

obtaining ~
P ik
c - I i
JL
a .

i
1 og
k
+ C
j

or
p ik
y~ a i

.
i og
* k

q
y y *
111
l i k i j l

The right side of this last expression is independent of

k, and so
V
^ a. -
i
log p - log^ q, /. = y
ii
.
* ik ^ k
ii
. a '

where y is independent of k. Then

log qk * =
Z r
i
log p. - Y

I-

18
 .

or

(29) V - r

i=l
ik i
r .

1
+ +r

where r.
l
= —
a .
, and r is constant. Thus in this case g,
k
* is
l

proportional to a weighted geometric mean, with weights a., of

the probabilities p
ik

(c) Linear utility .

Yet another possibility is to equate utility with money

This case was treated in detail by Gale and Eisenberg (1959)

and so we will merely refer the reader to that interesting

art icle

(d) A counter-example .

Let us consider a two-horse, two-bettor situation. Bettor

1 is certain horse 1 will win whereas bettor 2 feels the race is

up for grabs. They have equal capital. Thus

p ll = Z p 12 - o

21 22

A = A = 1
2 2

Bettor 1's utility function does not much matter - he will

bet all his money on horse 1. Bettor 2 has a utility function

3
x if X
~ 2
U (x) =
2
3
-
x
X >
~ 2

It is easily seen that, if g < — , then bettor 2 will choose

3q
= -
3q l
1
21 22 —-=-

whereas, if g > — , then 2 will choose

19
 3q 3q
*21 ' ' ~ — 22

For q = q - -, we would have - < x < -, x

22
= 1 - x .
2 2J 2J

Since x = 1, x = whatever q is, we will then have

b *(g)
2
= 1 + —
J
*l
if q2 <
1
-

V<*> - 2 - —
3q 2
=
1
2
+
3q
-J-
i
lf
*1 >
1
J

I s
V*' -
1 if
^i = i •

For an equilibrium, we must have b *(q) = 2q (since

c = a + A = 2 ) . But, from the above, we see this holds only

if g = 1. But this leads to the undesirable discontinuity on

the boundary of the simplex, and we must conclude that there is

no equilibrium for this situation.

20
 . ,

APPENDIX

Concavity and Convexity

A function f from mm to m is said to be concave if, for

every x , y e mm , and < X < l

(30) f(Xx + (l-r)y) > Xf(x) + (l-\)f(y).

It is strictly concave if strict inequality holds in (30) whenever
x * y and < X < 1 . A function F is convex "strictly convex]
if -f is concave [strictly concave] .

A set s c m is convex if, for any x,y e s and < X < 1

Xx + (l-X)y e S.

Generally, if f is a concave function, for any g, the set

• s = |
x | f(x) > q
q |

is convex . In particular, the set of all x which maximize f(x)

is convex (though it may be empty).

If a function f is strictly concave, it need not have a

maximum. If there is a maximum, however, the maximizing point is

unique

If f: m "R is concave, it will be di ff erent iable almost

everywhere in its domain. Even when not different iable , however,

f has both right and left derivatives, f'(x+) and f'(x-). The

derivative is monotone non-increasing, satisfying

31(a) f x+)( < f'(x-) for all x

31(b) f (x-) < f'(y+) if y < x.

If f is strictly concave, its derivative is strictly monotone,

satisfying 31(a) and satisfying 31(b) with strict inequality.

1,t
Upper Semi-Continuity

Let x, Y be topological spaces. A set-valued mapping from

X to Y is a mapping <j> which assigns, to each x e x, a subset

<$>(x) c y. It is a correspondence if <j> (x) * p for all x e x.

The set-valued mapping <J>

is said to be upper semi-continuous

if, whenever x -x* , y e (x ;, and y -y* then y*

* e ACx*;.
r
n ix
<J>
n n

Theorem . Let f be a continuous real-valued function defined

on the product space x x y. Define for x e x,

$(X) = /y | f(x,y) = max f(x,t)\

y tey '

then <J>
is an upper semi-continuous set-valued mapping from x to Y

If y is compact and non-empty, then <J>

is also a correspondence.

Kakutani's Fixed-Point Theorem .

Let x be a simplex in /R , and let <j> be an upper semi-

continuous correspondence from x to x, such that, for all x, $(x)

is compact and convex. Then there is some x* e x such that

x* e 4> ( x* ) .
 BIBLIOGRAPHY

1. Eisenberg, E. and D. Gale.

, "Consensus of Individual
Probabilities: The Pari-Mutuel Approach". Annals. Math. Stat
1959, 165-16S.

2. Kakutani, S. "A Generalization of Brouvver ' s Fixed Point

Theorem". Duke Math. J. 1941, 457-458.

3. Rockafellar, R.T. Convex Analysis . Princeton University

Press, 1970.

^
 DISTRIBUTION LIST

DEFENSE TECHNICAL INFORMATION CENTER (2)

CAMERON STATION
ALEXANDRIA, VIRGINIA 22214

CHIEF OF NAVAL RESEARCH (2)

ARLINGTON, VA 22217

LIBRARY, Code 0142 (2)

NAVAL POSTGRADUATE SCHOOL
MONTEREY, CA 93943

RESEARCH ADMINISTRATION (1)

Code 012
NAVAL POSTGRADUATE SCHOOL
MONTEREY, CALIFORNIA 93943

PROFESSOR GUILLERMO OWEN (8)

Code 530n
DEPARTMENT OF MATHEMATICS
NAVAL POSTGRADUATE SCHOOL
MONTEREY, CALIFORNIA 93943

iA
 DUDLEY KNOX LIBRARY

3 2768 00347544 3