BOOK REVIEWS

Theory of games and economic behavior. By John von Neumann and

Oskar Morgenstern. Princeton University Press, 1944. 18 + 625 pp.
$10.00.
Posterity may regard this book as one of the major scientific
achievements of the first half of the twentieth century. This will
undoubtedly be the case if the authors have succeeded in establishing
a new exact science—the science of economics. The foundation which
they have laid is extremely promising. Since both mathematicians
and economists will be needed for the further development of the
theory it is in order to comment on the background necessary for
reading the book. The mathematics required beyond algebra and
analytic geometry is developed in the book. On the other hand the
non-mathematically trained reader will be called upon to exercise a
high degree of patience if he is to comprehend the theory. The mathe-
matically trained reader will find the reasoning stimulating and
challenging. As to economics, a limited background is sufficient.
The authors observe t h a t the give-and-take of business has many
of the aspects of a game and they make an extensive study of the
strategy of games with this similarity in mind (hence the title of this
book). In the game of life the stakes are not necessarily monetary;
they may be merely utilities. In discussing utilities the authors find
it advisable to replace the questionable marginal utility theory by a
new theory which is more suitable to their analysis. They note that
in the game of life as well as in social games the players are frequently
called upon to choose between alternatives to which probabilities
rather than certainties are attached. The authors show that if a
player can always arrange such fortuitous alternatives in the order of his
preferences, then it is possible to assign to each alternative a number or
numerical utility expressing the degree of the player's preference f or that
alternative. The assignment is not unique but two such assignments
must be related by a linear transformation.
The concept of a game is formalized by a set of postulates. Even
the status of information of each player on each move is accounted
for and is characterized by a partition of a certain set. The amount
which player k receives at the conclusion of the play is a function
$k(cTi, C2, • • • , <rv) of the moves <ri, 0*2, • • • , <rv where some of the cr's may
be the moves of chance (dealing cards, throwing dice, and so on).
The concept of a game admits of a rather drastic simplification
which practically relieves the players of the necessity of playing.
498
 BOOK REVIEWS 499

Imagine that all possible strategies of all players have been cata-
logued. Then player k can tell his secretary that he wishes to play
strategy r^. When she looks up this strategy she finds a complete
prescription determining every move for every possible eventuality.
Thus the secretaries could get together and determine the outcome
of the game if they could only find an equitable method of accounting
for the moves of chance. But chance enters into the game very much
as one of the players. Thus we can imagine a cataloguing of the
possible strategies of chance. Suppose for the moment that the
strategy T 0 of chance has been decided upon and that the players
have chosen respectively the strategies Ti, T2, • • • , r n . Then the
strategies determine the moves. Hence fÇ^Ccri, CT2, • • • , <rv) is a function
®fc(ro, TI, • • • , r») of the strategies and the outcome of the game is de-
termined. But how should To be selected? Instead of selecting r 0 the
secretaries could assign to each player k the amount $fc(ri, T2, • • • , r n )
which he would receive on the average if strategies n , T2, • • • , rn of the
players were chosen. The amount $£h is the mathematical expectation
of ®fc. I t is computed in terms of the probabilities of the various
strategies r 0 and these probabilities are in turn computed in terms of
the probabilities of the moves of chance.
The game has now been reduced to one in which each player makes
just one move—the selection of a strategy. Each player makes his move
in complete ignorance of the moves of the other players. The authors
have accomplished this simplification of the game with complete
rigor and with complete adherence to the rules laid down by the
postulates.
A 1-player game corresponds to the economy of a man on a desert
island. I t is the Robinson Crusoe economy or a strictly regimented
communism. If the player is wise, he will choose his strategy n so
that Jpi(ri) is a maximum. This is the only case where a game is
settled by simple maximum considerations.
An w-player zero-sum game is one for which the sum of the $kS is
zero for all choices of <ri, 0*2, • • • , <r„ and hence one for which the
sum of the ^p&'s is zero for all choices of n , T2, • • • , r n . Social games
are zero-sum but the game of economics is decidedly not zero-sum
since society as a whole can improve its status if all members behave
properly. However an arbitrary n-player game can be reduced to a zero-
sum (n+l)-player game by introducing a fictitious player n+1 who
receives the amount ^pn+i(Ti, T2, • • • , TW) which is the negative of the
total received by the remaining n players. Note that the functions fQk
do not contain the variable rw+i, that is, the fictitious player is not
permitted to choose a strategy. It will appear later that further re-
500 BOOK REVIEWS (July

strictions on the activities of this player are necessary to prevent his

influencing the outcome of the game.
A zero-sum 2-player game T can be characterized by a single func-
tion § ( r i , T2) = ^ i ( r i , T2) since § 2 ( T I , r%) = — §( T i» T%) by virtue of the
relation § 1 + ^ 2 = 0. In this game player 1 will attempt to maximize
§ (or §1) whereas player 2 will attempt to minimize & (or maximize
§2). Since these are diametrically opposed tendencies, it looks as
though nothing could be decided. However we can gain insight into
the problem by considering a modified game Ti which is the same as
r except that player 1 moves first and player 2 knows l's move. In
Ti after player 1 chooses n , player 2 will choose T2 so as to minimize § .
It is therefore advisable for 1 to choose T\ SO as to maximize
minT2 ^ ( T I , T2) where minT2 &(ji, ^2) is the minimum with respect to
T2 of Jp(ri, T2). Player 1 will then receive.
Vi = max m i n § ( r i , T2)

and 2 will receive —»i. Next consider a third game T2 which is the
same as V except that 2 moves first and 1 knows 2's move. If both
players of T2 are skillful, then 1 will receive the amount
V2 = min m a x ^ ( n , r 2 )
T2 n

and 2 will receive — z>2. In the original game T if both players are
skillful, 1 will receive at least V\ and at most V2 whereas 2 will receive
at least —vi and at most —Vi. Hence Vi^vz and these quantities are
bounds f or the outcome of the game. If Vi — v*, the game is determined
but in general this is not the case.
Note that T reduces to Y\ if 2 discovers l's strategy whereas T
becomes T2 if 1 discovers 2's strategy. Hence it is advisable for the
players to conceal their strategies. The concealment is accomplished
by using probabilities. Thus 1 chooses T\ with probability £T1 and 2
chooses T2 with probability rjTi. The average outcome J£(£, 77) for player
1 is the mathematical expectation of § ( r i , T2) with respect to the prob-
abilities £T1 and rjT^ where £ is the vector with components £1, £2, • • •
and rj is the vector with components 771, 772, • • • . The introduction of
these probabilities modifies V and consequently modifies Ti, I \ and
the bounds vi, V2. The new bounds become
v( = min max 2£(£, rf)
and
vi = min max ÜT(£, tj).
19451 BOOK REVIEWS 501

It is easily shown that Vi^v{ ^vi ^%, that is, that each player is at
least as well off as before the probabilities were introduced. Moreover
it can be shown that
v( = vi = v
and hence that the game is determined. The proof of the latter result
depends on the fact that the numbers #T2= ]CTI§( T I» r*)%n a r e com-
ponents of a vector % which depends on £ and that the tips of the
vectors x for all possible £*s constitute a convex set of points.
Next consider an w-player game in which the players divide into
two hostile groups called S and — 5. This can be interpreted as a
2-player game between the players S and —5. If probabilities are
employed in the manner described above, then 5 will receive
v(S) = v{ = vi = v
and —5 will receive
, ( - S) « - v(S).
If I is the set of all players, then z/(J) = 0, that is, the game is zero-
sum. Finally
v(S + T) £ v(S) + v(T)
if S and T are mutually exclusive groups. That is, the players of
S+T can obtain at least as much by cooperating as they can by
splitting up into two groups S and T. The function v(S) satisfying the
above relations is called a characteristic function. Corresponding to
any function satisfying these relations there exists a game having
this v(S) as its characteristic function. The construction of such a
game involves partitions of I into subsets called rings and solo sets.
If the equality v(S+T)=v(S)+v(T) always holds, that is, if v(S)
is additive, then the coalitions will be ineffective and the game will
be determined. This is the case for n = 2. Moreover two characteristic
functions (whether additive or not) which differ by an additive func-
tion will produce the same strategies of coalitions. If v(S) is not addi-
tive, it can be modified by a suitable additive function and a suitable
scale factor so that v(S)~ —1 for all 1-element sets. Thus for n = 3,
v(S) is given by the following table
0 10-element set (— I or the complement of I)
—1 ) 1-element sets
v(S) for the \
+ 1 / 2-element sets (complements of Vêlement sets)
0 ^S-element set (/).
502 BOOK REVIEWS [July

For w ^ 4 , v(S) is no longer determined and the number of possibil-

ities becomes almost bewildering. The reader will begin to realize that
there is never a dull moment with these games. We have seen that
for each of the cases » = 1, 2, 3, 4 a new situation appears. For n = 5
no new phenomenon has as yet been discovered but for n <£ 6 we first
meet the possibility of a game which splits into two or more games
which are in some respects quite distinct but which nevertheless exert
potent influences on one another. This phenomenon has the counter-
part of nations whose economies are distinct yet interdependent.
It remains to consider what coalitions can be expected to form in
a given game and how the stakes will be divided in the presence of
such coalitions. A division of stakes is called an imputation and is
represented by a vector a with components «i, a% • • • , an where a&
is the amount the feth player receives. One could imagine that if a
group of novices were playing one of these games a certain chaos
would result. Coalitions would be made and broken as each player
sought to improve his own status. Finally as the players became
more acquainted with the game certain imputations would come to
be trusted because of the stability of the corresponding coalitions and
because of the profitableness to an effective group of players. There
would thus emerge a set V of trusted imputations. There would of
course be players who were dissatisfied with any given trusted im-
putation but they would not be strong enough to force a change
unless they could bribe some of the favored players to desert their
coalitions. Nor would such bribery be effective since the potential
recipient of the bribe would realize that the chaos produced by his
desertion would eventually leave him in a less favorable position.
Thus V corresponds to a group behavior pattern. I t is an institution
or a morality arising from enlightened self interest.
But how can V be described mathematically? We begin with a
definition. We say that an imputation a dominates an imputation j8 if
there is an effective group of players each of which is better off under a
than under j8. The group is effective provided it can guarantee for its
members the stakes prescribed by a against any opposition from without
the group. A set V of imputations is called a solution provided every
imputation outside of V is dominated by some imputation of V and
no imputation in V is dominated by any other imputation in V. Thus
V is a maximal set of mutually undominated imputations. Unfor-
tunately dominance does not produce even a partial ordering of the
set of all imputations. I t is not a transitive relation. This makes the
discovery of solutions a difficult task. We shall however outline a
method of finding solutions for the case w = 3.
19451 BOOK REVIEWS 503

When « = 3 w e have
a — («i, «2, 0:3) with a i + «2 + «3 = 0,

t h a t is, the game is zero-sum. Thus the tip of a lies in a plane which
passes through the origin and is equally inclined to the coordinate
axes. This plane is divided into six congruent sectors by the traces of
the coordinate planes. Next a^ — 1 (for k = 1, 2, 3) since each player
can obtain at least — 1 without the benefit of any coalition (see the
above table). These inequalities require the tip of a to lie within an
equilateral triangle whose center is at the common intersection of the
traces of the coordinate planes and whose sides are parallel to these
traces. An imputation a dominates those imputations which are rep-
resented by points interior to three parallelograms each of which has
two sides in common with the above equilateral triangle and one
vertex a t the tip of a. On the basis of these geometrical considerations
it is easy to find solutions V, We first look for a V whose imputations
do not all lie on a line a& = a constant (that is, a line parallel to a
trace). There is only one such solution, namely,
V: (1/2, 1/2, 0), (1/2, 0, 1/2), (0, 1/2, 1/2).
We next look for a V whose imputations do lie on a line, say, <xz~c.
The corresponding solutions are
Vc: (a, — a — c, c)
where a and c are required to satisfy certain inequalities. Thus Vc
contains a continuum of solutions corresponding to values of the
parameter a. This exhausts the possible solutions. The first solution
V seems quite reasonable whereas Vc seems unnatural and difficult to
interpret b u t let us return to this question later.
Let us consider the following non-zero-sum 2-player game. Each
player (1 or 2) chooses either the number 1 or the number 2. If both
players choose 1, then each receives the stake 1/2. Otherwise each
receives — 1 . If we reduce this game to a zero-sum 3-player game by
the introduction of a fictitious player 3, then the characteristic func-
tion becomes the one given in the above table. Now if we take the
first solution F, we discover t h a t the fictitious player may play an
active part in the formation of coalitions. Hence if we wish to retain
the 2-player character of the game, we must choose the solution Vc
and it is reasonable to assign to c the value — 1 .
The authors apply this theory of games to the analysis of a market
consisting of one buyer and one seller and also of a market consisting
of two buyers and one seller.
504 BOOK REVIEWS [July

The book leaves much to be done but this fact only enhances its
interest. I t should be productive of many extensions along the lines
of economic interpretation as well as of mathematical research. In
fact the authors suggest a number of directions in which research
might profitably be pursued.
ARTHUR H. COPELAND

Principles of stellar dynamics. By S. Chandrasekhar. The University

of Chicago Press, 1942. 10+251 pp.
The primary field of this book is astronomy and not mathematics,
although the latter is used as an essential tool. The readers of this
review, professional mathematicians almost exclusively, will have a
normal human interest in the major astronomical aspects of the book,
but their critical scrutiny is bound to be concentrated on how the
astronomical problems are formulated mathematically and what sort
of mathematics has been proposed for their solution. For this reason,
and partly also in the interest of brevity, this review treats only of the
mathematical aspects of the book.
In the first chapter is given a detailed discussion of the kinemati-
cal concepts appropriate to the study of stellar systems. Since these
systems contain a large number of stars, it becomes necessary to intro-
duce a method similar to that employed in hydrodynamics, where the
motion of a fluid is described by a vector field, representing at each
point and for each instant of time the velocity of the fluid. In hydro-
dynamics the velocity of the fluid at a point is conceived as the veloc-
ity of the "fluid particle" a t the point in question. But this notion of a
"particle" at the point in question is difficult to make precise, espe-
cially if one assumes the fluid to consist of a large number of small
atoms with relatively large empty spaces between them. Nevertheless
such a concept (in which the stars play the role of the atoms) is char-
acteristic of stellar dynamics as distinguished from celestial (particle)
mechanics, which considers systems containing but a relatively small
number of bodies.
The components U0(x, yy z, /), Fo(#, y, z, /), WQ(X, y, s, t) of the vec-
tor field thus introduced do not, of course, necessarily represent the
components of velocity of a star which might happen to be a t the
point (x, yy s), but rather the velocity of the centroid of stars in a
"small volume" about the point (x, y, z). The components of velocity
of an individual star are written in the form Î7= Uo+u, V= VQ+V,
W= Wo+w, where the vector (u, v, w) is called the residual velocity.
The statistical consideration of these residual velocities is a charac-
teristic of stellar dynamics and gas theory as distinguished from