PROBABILITY

A HISTORY OF PROBABILITY IN THE UNITED STATES OF

AMERICA BEFORE THE TWENTIETH CENTURY
WILLIAM DOWELL BATEN

This article contains a history of the material concerning probability which

was written by people who lived in the United States of America. I t begins with
an article presented in 1791 and extends through 1900. During this period many
papers on probability were presented to the various scientific organizations.
The subject matter includes life insurance theory, mortality tables, derivations
of the normal curve of error, the method of least squares, probable errors of
observations, games of chance, the importance of probability problems, prob-
ability laws of functions of chance variables, derivations of various probability
laws, inverse probability, properties of polynomials, mean values, gambler's
ruin, geometric probability, test books, teaching concerning probability, applica-
tions, probability courses, etc.
This paper also lists the names of the people who presented this material and
summarizes their contributions to the development of probability subject matter.
MICHIGAN STATE COLLEGE,
EAST LANSING, MICH., U. S. A.

ON ASYMPTOTIC EXPANSIONS OF PROBABILITY

FUNCTIONS
HARALD BERGSTRöM

Consider the pr. f. of a sum Z ( 1 ) + Z ( 2 ) + • • • + Xin) of n independent,

equally distributed random variables in the euclidean space Rk of the finite
dhnension k. If each variable has the pr. f. P(E), the sum has the pr. f. P*n(E),
where P*n(E) denotes the n-fold convolution of P with itself. In particular for
1-dimensional d.f.'s the known Edgeworth expansion has been used as an ap-
proximation of the d.f. for the sum. In the following we give a more general
asymptotic expansion of P* n from which the Edgeworth expansion may be
determined as a special case.
If <p (E) is a set function and E a set, we denote the total variation of cp in
Rk by V[<p] and put Ms[<p] = Max* | <p(E — x) \ , where E — x denotes the set
obtained by E through the translation —x. g(n) (as well as gi(n, s)) may denote
a decreasing function of n with g(n/p) = 0[g(n)] for every positive constant p.
Let now Q be a pr. f. which shall be used for the approximation of P* n . Then

(Ai) P*n = È A?} + ri8+1)

565
566 SECTION IV. PROBABILITY AND STATISTICS

with A«*0 = Cn>vQ*w""*(P - Q)*' and some remainder term r(ns+1). If the con-
dition
(CO V[A?] < g{n)
is satisfied, A»} = 0[gv(n)] for any fixed integer v. However A ^ may be of a
smaller order of magnitude than gv(n). Let now (Ci) hold and let S =
X)n-i gs+1(n)M converge when s is some fixed integer ^ 0 , and suppose that
the inequalities ME[A{ns+1)] < v(E)gi(n, s), ME(P) < p(E), MB(Q) < p(E) and
the condition
(CO Hm V[P*P*(P - Q)] = 0
are satisfied. Then
(1) ME[ris+1)l < Cß(E)gi(n, s),
where C depends on s and the functions g(n) and g±(n, s) but not on n. When
(Ci) and (C2) are satisfied and S converges, then (Ai) has moreover the strictly
asymptotical properties
(2) 7[A„"] = _ % » ] , V[ri'+1)] = 0\g'+\n))
and (Ci) and (C2) are necessary for the existence of these properties when S
converges. The condition (C2) is used only to that extent that V[P*n*(P — Q)]
is smaller then some value depending on s and the functions g(n) and gi(n, s)
but not on n.
When P is a nonsingular d.f. with finite absolute mean values of the order X
for some X in the interval 2 < X g 3, and Q is that (nonsingular) normal d i .
which has the same mean value vector and the same moment matrix as P, then
(Ci) is satisfied with
g(n)=n-*-2Wlogwn
and Air) = o(«~'<x_2)/2) or 0[n""iii~2W] according as X < 3 or X = 3. We have
(3) ri*» = 0[n~w + n-(s+1>(X-2)'2]
always, and
(4) ri8+1> = 0[«-(8+1)(X-2)/2]
when (C2) holds. Here (C2) may be replaced by the less restrictive condition
(C3) lim V{P*n{x)*Q[(an)"x]*[P(x) - Q{x)]\ = 0
TO—*00

with a = (s + 1)(X — 2)/2 — 1/2 and a constant a > 1. In the 1-dimensional

case, (C3) is necessarily satisfied if
ri+» = o(rr (5+1)(X - 2)/2 ).
From (Ai) different expansions may be derived. Let P be a nonsingular pr. f.
with finite absolute mean values of order X for some X in the interval 2 < X ^ 3,
 PROBABILITY 567

and let Q be that normal pr. f. which has the same mean value vector and the
same moment matrix as P. Further suppose that ME(P)/m(E) is bounded,
when E belongs to any family, regular in the sense of Lebesgue, and m(E) > 0
is the measure of E. Then denoting the derivative in the sense of Lebesgue by
D[<p] for a set function (p, we have
(5) D[A[V)] = 0[n-(*+KX~2>)/2]
everywhere and, if the condition (C2) is satisfied,
(6) D[ri'+1)] = 0[«-<t+(*+1,<x-2))/2]
in all points, where DP*n exists. If P is absolutely continuous, the condition (C2)
is necessary for the validity almost everywhere of that last relation.
CHALMERS UNIVERSITY OF TECHNOLOGY,

GOTHENBURG, SWEDEN.

RANDOM FUNCTIONS ON DIVISIA ENSEMBLE

MARIA CASTELLANI

In this paper we investigate a class of real measurable functions of a statis-

tical Divisia ensemble. Let us consider a lattice space with a set D of elements
which renew themselves at random during a certain time interval (tQ, t), be-
cause there is an inflow and outflow of elements. Let us assume that each ele-
ment is associated with a real, measurable, bounded function 8(6 — x) for
(6 - x) C (t - *o).
Let p(6 — x) be the probability function of an element which is not being
eliminated from D, dye the probability density of an element which is eliminated,
dvx the elements which areflowinginto D during the time interval (x, x + dx),
and Xo the elements in D at the time to. In several problems it might be more
convenient to assume p(6 — x)-8(9 — x) = f(B — x). According to any given
law of random flow, the expected value E (8) is

E(8) = Xo/(* - Q + Xo f f(6 - k) dße

JtQ
pt ft pt
+ /(. - x) dvx+ / f(6 - x) foe dvx.
J to J to Jx
We do not investigate this general integral equation but, instead, limit our
attention to obtaining some fundamental equations of actuarial mathematics.
When the ô is a linear function and we assume a Karup functional relation
p(6 — x) dße = — dp (9 — x), we then obtain a generalized Galbrun integral
equation for sickness insurance. A generalized Cantelli equation for mathe-
matical reserves is easily derived by considering the actual rate of elements
flowing out.
The Benini-Des Essars equation for the average time spent in D by the flow-
568 SECTION IV. PROBABILITY AND STATISTICS

ing elements is derived under very restrictive conditions of continuity and

linearity.
UNIVERSITY OP KANSAS CITY,
KANSAS CITY, M O . , U. S. A.

AN ERGODIC THEOREM FOR STATIONARY MARKOV CHAINS

WITH A COUNTABLE NUMBER OF STATES
KAI LAI CHUNG

Let Xn, n = 0, 1, • • • , be the random variables of a Markov chain with a

countable number of states, numbered 1, 2, • • • , and with stationary transition
probabilities P$. As usual P^ denotes the n-step transition probability.
Write Tffi = YL^P^fk- The ergodic theorem for Markov chains states that
if j and k belong to the same ergodic class, then lim^-«,n^T^ = p* - I t is
known that if p3 > 0, then pk > 0 so that the above implies lim».*,* Tff/Tif =
Pj/pk . If Pi = 0, then pk = 0 and no conclusion can be drawn of the exist-
ence of the last-written limit. This question was raised by Kolmogorov and
answered by Doeblin (Bull. Soc. Math. France vol. 66 (1938) pp. 210-220). More
precisely he proved that if j and k are such that there exist positive integers m
and m' for which P g ° > 0 and P f r 0 > 0, then l i m ^ Ttf/Ttf = 1,
and lim^oo T^/Tik^ exists and is positive and finite.
A new proof of Doeblin's theorem is given by introducing the quantities in
the infinite series written below and using generating functions. This method
also determines the last-written limit to be equal to X3/Xk where Xy = 1 +
S n=i P (Xv ^ k for 1 g v ^ n, Xn = j \ X0 = j) and XÄ is obtained by inter-
changing j and k in Xy. With extraneous conditions on the analytic nature of
Tik) the theorem becomes a consequence of a Hardy-Littlewood-Karamata-
Tauberian theorem. Our proof circumvents this invocation. An extension is
made to Markov chains with a continuous parameter t. If e. g. the uniformity
condition lim* ; 0 Pj3(t) = 1 uniformly in all j is assumed, Doeblin (Skandinavisk
Aktuarietidskrift vol. 22 (1939) pp. 211-222) has shown that almost all sample
functions are step functions, etc. Under these circumstances it is proved that
limi-oo JQP3J(S) ds/}oPkk(s) ds exists and is positive and finite, in fact a deter-
mination of its value similar to the above is given. Contrary to hope this na-
tural extension is not easily deducible from the discrete parameter case, but
our method works with Laplace transforms instead of generating functions.
The question of the existence of lim^«, Pff/Pi^ (or of lim*-«, Pjj(t)/Pkk(t))
seems to be open. Only in the special case where Xn is the sum of n independent,
identically distributed random variables with mean 0, it has been proved by
Erdös and the present author that this limit is one.
CORNELL UNIVERSITY,
ITHACA, N. Y., U. S. A,
 PROBABILITY 569

ELEMENTS ALÉATOIRES DE NATURE QUELCONQUE

R. FORTET

 la suite de travaux de M. Fréchet, je me suis intéressé, et j'ai incité quel-

ques jeunes probabilistes à s'intéresser, à l'élaboration d'une théorie directe
des éléments aléatoires de nature quelconque (ne se réduisant pas nécessaire-
ment à des variables aléatoires numériques à une ou à un nombre fini de dimen-
sions) ; des applications à la théorie des fonctions aléatoires ou à divers problèmes
concrets justifient cette étude; il est naturel, surtout si on a en vue l'application
aux fonctions aléatoires, d'aborder d'abord le cas d'un élément X prenant ses
valeurs dans un espace linéaire 36; un premier problème est alors d'étendre au
cas d'un tel élément la notion d'espérance mathématique, ce qui revient à une
théorie de l'intégration; dans le cas où 36 est un espace de Banach, Mlle Mourier
a montré (CR. Acad. Sci. Paris t. 229 (1949) p. 1300) qu'une application de
l'intégrale de Pettis permet de perfectionner notablement des résultats de
Fréchet; tandis qu'une extension, non encore publiée, de la méthode de Danieli
faite par M. Régnier permet de définir une espérance mathématique pour une
large catégorie de f.a. mesurables au sens de Doob.
Un second problème est d'envisager l'addition d'éléments aléatoires X i ,
X2, • • • , Xn, • • • à valeurs dans 36; pour cela la notion de caractéristique
semble devoir être utile; si 36 est un espace de Banach, on peut définir la carac-
téristique <p(x*) d'un élément aléatoire X, comme fonction de la fonctionnelle
linéaire x* variable dans le dual 3£* de 36, par <p(x*) = E[x*(X)]\ des propriétés
de <p(x*) s'établissent aisément, dont la principale est que <p(x*) est "définie
positive"; mais une fonction définie positive <p(x*), même continue avec la
topologie faible dans #*, n'est pas forcément une caractéristique; si 36 est sepa-
rable et réflexif, on peut donner une condition nécessaire et suffisante pour que
<p(x*), supposée définie positive, soit la caractéristique d'un élément aléatoire
X "proprement dit" (c'est-à-dire tel que Pr [ || X || < +°o] = 1); on peut
définir un élément laplacien X par la condition que, en supposant E(X) =
6 (0 étant l'élément 0 de 36), le logarithme de la caractéristique <p(x*) est de
la forme: — (l/2)E[x*(Z)], où Z est un élément aléatoire proprement dit quel-
conque à valeurs dans 36.
UNIVERSITY OF CAEN,
CAEN, FRANCE.

A NOTE ON THE GENERAL CHEBYCHEFF INEQUALITY

SALEM H. KHAMIS
a x
Let $(x), = = fr> be a cumulative distribution function, differentiable,
never-decreasing and possessing moments up to order 2n. Let §(x) be defined
as 0 for x S o, and 1 for x ^ b, and assume that it has at least n + 1 points
of increase. If ^?(x), a! ^ x g V, be another cumulative distribution function
570 SECTION IV. PROBABILITY AND STATISTICS

with similar properties and identical moments up to order 2n, then, by the
Chebycheff inequality, |$(s) - ^(x) | ^ [Q»(aOQ»+i(x) - QnfaOQn+ifc)]-1 =
fi2n(#), say, ^here, Qr(x), r = 0, 1, 2, • • • , or n, is the denominator of the rth
convergent of the continued fraction associated with ftZ d$(t)/(x — t) or
/ Ì * d&(t)/(x — t) and is a polynomial of exact degree r possessing orthogonal
properties with respect to d$(x)/dx and d$f(x)/dx as weight functions.
We have shown that the right-hand side of the above inequality may be
improved upon, obtaining for it the value Ati2n(x), where A is given by

0^4 = 1 + mm max < — ^JT-T-}, max < — ^-rr-T-f è 1,

provided the maxima are finite.

We have shown also that when the property "never-decreasing" does not
hold, the difficulty, under certain general conditions, could be overcome, thus
making it possible to apply the Chebycheff inequality to approximations for
distribution functions by series expansions involving negative components.
Further, the general relationship between the polynomials Qr(x) and the corre-
sponding orthogonal polynomials, as usually defined in treatise on orthogonal
polynomials, is given.
The methods used in this paper are illustrated, particularly, in the case of
the Pearson type III distribution function.

STATISTICAL OFFICE, UNITED NATIONS,

LAKE SUCCESS, N. Y., U. S. A.

UNLIKELY EVENTS IN GENERAL STATIONARY-

TRANSITION MARKOFF CHAINS

BERNARD OSGOOD KOOPMAN

A system is considered whose possible states form a perfectly general ab-

stract set X, the observable subsets of which form a cr-algebra 2 . It evolves
stochastically through a sequence of n stages, at the fcth of which it is in the
state xk € X (k = 1, • • • , n). The transition probabilities pntk(%, E) =
probfe Ç E | Xk~i — x) are one-step Markoffian (x £ X, E £ 2). For a certain
S Ç 2, the event [xk £ S\ is called "success on the kth trial". The object is to
find P(s) = limn_>00Pri(s), where Pn(s) = probability of s successes in first n
trials.
It is shown that the class S3 of all complex-valued functions / = f(x, E), de-
fined and bounded over (x, E) £ (X, 2), tr-additive in E (x fixed) and meas-
urable (2) in x (E fixed), is a Banach algebra, provided linear combinations
with complex coefficients are defined as usual, the SB-product fg = (fg)(x, E)
as the abstract integral of g(y, E) with respect to the measure f(x, Fv)((y, Fy)
 PROBABILITY 571

being variables of integration), and norm N(/) as supxeX [absolute variation

of f(x, E) on X]. The unit element u £ S3, where u(x, E) = XE(X) = characteris-
tic function of set E. Accents and 0 denote complements and intersections in X;
and the ordinary (non 93-) product is written f(x, E)g(x, E).
The following hypotheses involve the stationarity of transitions, the im-
probability of success, and the Markoffian regularity of hmm^w hm = g in (4) :
(1) Success ~> success: xa(x)pn,k(x) Sd E) == a(x, E), N(a) < 1.
(2) Success -> failure: xs(x)Pn,k(%, S' 0 E) = b(x, E).
(3) Failure —> success: Xa'(x)Pn,k(x) 8 Ci E) = c(x, E)/n.
(4) Failure —> failure: Writing xs>(x)pn,k(x> S' H E) as gn,k(x7 E)Pn,k(x> S'),
where gn,k(x> E) = Woh(xk 6 E \ xk-i = x, xh G S'), we assume: (i) gn%k(x, E) =
g(x, E)\ (ii) g(x, E) è xs>(x)v(E), where v(E) is c-additive over 2, v(E) è 0,
v(8') > 0, v(Sf fi E) = ?(_E). By classical reasoning 0m - * A, where ä(&, J®) =
X8'(x)v(E)9 and ju(#) is like K#), but /x(S') = 1.
Then it is shown that P(s) is generated by the function <p(t) = ]C?=o P(s)t8
obtained by first constructing the Banach-algebraic expression
ft = Po[(u — at)^b + u] g exp [(t — l)c(u — at^h],
where h(x, E) = xs(x)v(E), and pQ(x, E) = po(E) is the initial probability that
Xi G E] and then setting <p(t) = ft(x, X) (which is independent of x).
An equally explicit but more complicated result is obtained in the non-
stationary case.
COLUMBIA UNIVERSITY,
N E W YORK, N. Y., U. S. A.

SOME NON-NEGATIVE TRIGONOMETRIC POLYNOMIALS

CONNECTED WITH A PROBLEM IN PROBABILITY
EUGENE LUKACS AND OTTO Szlsz
Let 0 < 6i < b2 < • • • < bn be n integers and let g(6) be the Vandermonde
determinant formed from b\, b\, • • • , b2n with the first row replaced by
sin2 bid/2 (i = 1, • • • , n). The function g(6) is then a cosine polynomial. In con-
nection with a problem in probability the question arose as to when g (6) is a
non-negative trigonometric polynomial. The following results were obtained:
(A) If the bi are the first n consecutive integers, then g(0) is non-negative.
(B) If the bi are the first n consecutive odd integers, then g(6) is non-negative.
(C) If the numbers &i, • • • , bn are obtained from the first (n + 1) consecutive
integers by omitting the integer k (1 ^ k ^ n), then the trigonometric poly-
nomial g (6) is non-negative if and only if 2fc2 ^ w + 1 .
(D) If the numbers bi, • • • ,bn are obtained from the first (n + 1) consecutive
572 SECTION IV. PROBABILITY AND STATISTICS

odd integers by omitting the kth odd integer 2k — 1 (1 ^ k ^ n), then the
trigonometric polynomial g (6) is non-negative if
1/2
i k/ ^•> Äi
1. i. where
i, 7 = 1 + (1
ßi ^—+ 2n)— or if.,
rl /-, in-l/,2 2\
2.k <h but if fLXQ (1 - f ) n _ 1 (r - xl) dt^O where

/ n + 2fc - 2fc2 \1/2

*° \ > ( n + 1) - 2k(k - 1 ) / '
If neither of these conditions is satisfied, g(6) assumes also negative values.
(E) If the numbers &i, • • • , bn , are chosen from the first (n + 2) consecutive
integers by omitting two integers k and p (1 ^ k ^ p — 1 g n), the non-nega-
tiveness depends on the discussion of a quadratic polynomial whose coefficients
are functions of n, k, and p.
NATIONAL BUREAU OF STANDARDS,
WASHINGTON, D. C , U. S. A.
NATIONAL BUREAU OF STANDARDS,
Los ANGELES, CALIF., U. S. A.
UNIVERSITY OF CINCINNATI,
CINCINNATI, OHIO, U. S. A.

INFORMATION RETRIEVAL VIEWED AS TEMPORAL

SIGNALLING

CALVIN N. MOOERS

The problem of directing a user to stored information, some of which may be

unknown to him, is the problem of "information retrieval". Signalling theories
can be applied, though this is a form of temporal signalling, which distinguishes
it from the point-to-point signalling currently under study by others.
In information retrieval, the addressee or receiver rather than the sender is
the active party. Other differences are that communication is temporal from
one epoch to a later epoch in time, though possibly at the same point in space;
communication is in all cases unidirectional; the sender cannot know the partic-
ular message that will be of later use to the receiver and must send all possible
messages; the message is digitally representable; a "channel" is the physical
document left in storage which contains the message; and there is no channel
noise because all messages are presumed to be completely accessible to the re-
ceiver. The technical goal is finding in minimum time those messages of interest
to the receiver, where the receiver has available a selective device with a finite
digital scanning rate.
Classification and indexing schemes are ruled out because of gross topological
 PROBABILITY 573

difficulties, and the Batten system because of no reasonable large-scale me-

chanization. To avoid scanning all messages in entirety, each message is char-
acterized by N independently operating digital descriptive terms (representing
ideas) from a vocabulary V, and a selection is prescribed by a set of S terms.
Conventional assignment of a digital configuration to a message requires
N log2 V binary digit places. It is a nonsingular transformation from ideas to
configuration. Prescription of a selection requires S log2 V digit places. Although
only log2 M digit places are required to differentiate (or enumerate) M messages,
S log2 V may be many times greater, indicating a digital redundancy and waste
in coding. This redundancy can be removed by recoding the descriptive ideas
into the digital configuration representing the message by a singular transforma-
tion (Zatocoding). The message can then be characterized with (N/S) log2 M
digit places, with a possible increase in scanning rate per message. Because Zato-
coding represents each idea by a pattern ranging over the digit places for a
message, and superposes these in the same coordinate frame by Boolean addi-
tion, the selection process includes the desired messages and gives a statistical
exclusion of the undesired messages.

ZATOR COMPANY,
BOSTON, MASS., U. S. A.

ON A CLASS OF TWO-DIMENSIONAL MARKOV PROCESSES

MURRAY ROSENBLATT

Let x(i) be an element of Wiener space and let V(t, x) satisfy the local Holder
condition | V(t, x) - V(t + At, x + Ax) | g M(t, x) { \ At \a{tiX) + | Ax \ait'x) } ,
a(t, x) > 0 everywhere except on a curve rectifiable in every finite rectangle of the
(t, x) plane. Let V(t, x) be bounded in every finite rectangle of the (t, x) plane.
The study of the two-dimensional Markov process (x(t), jo V(r, x(r)) dr) is
related to the study of certain differential and integral equations. Let Q(t, x) =
E {e-"S!>VMT»dT | x(t) = x}e^2lu/(2irtf12. If 0 ^ V(t, x), Q(t, x) is the solution of
the differential equation (l/2)d2Q/dx2 — dQ/dt — uV(t, x)Q — 0, subject to the
conditions Q(t, x) —> 0, x —» db°°, lim^ 0 /- e Q(t, x)dx = lim*_>0 /"«> Q(t, x)dx =
1 for all e > 0. Qx(t, x) is continuous in (t, x) for all (t, x) 5^ (0, 0). Let G(x, y,
t) = Pr{/o V(r, x(r))dr g y \ x(t) = x}. Then /J*x (E(a) - G(x, a, t)) da =
2/2 1/2
ß* '(27r0 til^e-™'1*™-*1*/^^ - r)) 1 / 2 7(r, *){£(£, y + X, r) - O f t ,
y, T)} d£ dr where E(y) = 1 if y ^ 0, and = 0 if y < 0. This last equation is
valid if one assumes that V(t, x) is bounded and Borei measurable.
CORNELL UNIVERSITY,
ITHACA. N. Y., U. S. A.
574 SECTION IV. PROBABILITY AND STATISTICS

DISTRIBUTION FOR T H E ORDINAL NUMBER OF

SIMULTANEOUS EVENTS WHICH LAST
DURING A FINITE TIME
HERMANN VON SCHELLING

If we know (1) that an event of a given class may occur with the probability
p per unit of time; (2) that, when it has happened, it will not occur again during
the next k units of time; (3) that such an event was observed at the nth moment;
we ask for the probability w(m;n, k, p) that this event is the rath one since zero
time.
Relative probabilities can be found immediately. Their sum sn(x) with x==
p/q*1 satisfies the following recursion formula

Sn+l(x) — Sn(x) — XSn-k(x) = 0

with Si(x) = s2(x) = • • • = sk+i(x) = 1. Its solution depends upon the roots
of the characteristic equation

/(*) = zk+1 - qzk - 1 + q = 0.

It has to be proved that (a) f(z) = 0 does not have multiple roots; (b) z\ = + 1
is the only root of f(z) = 0 on the unit circle; (c) the roots z2, z*, • • • , zk+i of
f(z) = 0 are located inside the unit circle.
Exact formulas for w(m\n, k, p) and for the first and the second moment of this
distribution are given as functions of the roots of f(z) = 0. According to (a),
(b), (c) it is possible to present simple approximations which do not depend on
the solution of the characteristic equation and may be evaluated easily.
The distributions w(m; n, k, p) are branching from the binomial type which
corresponds to the particular case k = 0. The final results are elementary. It
is to be expected that the new distributions will become useful in many fields
of science and biology since the assumption (2) is frequently adequate to ob-
served conditions.

NAVAL MEDICAL RESEARCH LABORATORY,

N E W LONDON, CONN., U. S. A.

INFORMATION AND T H E FORMAL SOLUTION OF

MANY-MOVED GAMES
L. S. SHAPLEY

A game is said to be in normalized form if it is so formulated that each player

has just one move—the selection of a strategy—made independently of the
other players' selections. A 2-person zero-sum game with n moves in which no
 PROBABILITY 575

information passes between moves can easily be formulated in this fashion, and
its solution is obtained by considering:

(1) max min Jj H(xi, • • • , xn) d<p(xh , • • • , xip) d$(x3l, • • • , xjn_p)

where tp is a probability distribution over the set of strategies (XìX , • • • , Xip)

of the maximizing player, and \j/ is the same for the minimizer. We suppose that
the identity of the player making the ith move, as well as the set Ai of his alter-
natives, does not depend on the previous course of play. We may abbreviate (1)
to
mix H(xi, ' • • , xn)
a\i-a
where I is the set of integers { 1, • • • , n } and S the set of moves assigned to the
maximizer.
The introduction of information (i.e. knowledge, at one move, of the outcome
of another) quickly makes the normalized form useless in practice, since the
number of strategies increases at a prohibitive rate. Yet, in the extreme case of
perfect information (knowledge, at each move, of all preceding moves), the
game is relatively easily solved one move at a time, e.g. :
(2) max min min • • « max H(xi, x2, x%, • • • , xn).
Xi£Ai W2ÇA2 *sE^3 «n£An

I t is of interest to ask what conditions on the informational structure are re-

quired for a game to be decomposable into a sequence of separated moves (as
in (2)), or, more generally, separated subgames in normalized form. The solution
in such a case would be given by:
(3) mix mix • • • mix H(xi, • • • , xn)
Ti\Ui T2\U2 Tm\Um

where the Tk and Uk partition S and I — S respectively. Using McKinsey's

notion of equivalence of patterns of information, we find that a necessary and
sufficient condition that a game be solvable by (3), regardless of the nature of
the function H or of the sets Ai, is that it be possible to order the moves in time
so that (i) the information pattern never implies knowledge of the future, (ii)
each player's knowledge of his opponent's actions increases monotonically, and
(iii) the accessions of information occur only between' certain pairs of consecutive
moves, on which occasions both players are apprised of all their opponent's
actions to date. No condition is imposed on a player's awareness of his own pre-
vious actions.

THE RAND CORPORATION,

SANTA MONICA, CALIF., U. S. A.
576 SECTION IV. PROBABILITY AND STATISTICS

STOCHASTIC PROCESSES BUILT FROM FLOWS

KÔSAKU YOSIDA

By virtue of the theory of semi-groups due to E. Hille and the author, we may
construct stochastic processes in a separable measure space R from flows in R.
A flow ( = one-parameter family of equi-measure transformations in R) Ftx
induces a one-parameter group Tt of transition operators: (Ttf)(x) = f(Ftx),
— oo < t < oo. Here a linear operator T on the Banach space Li(R) to L±(R)
is called a transition operator if f(x) à 0 implies (Tf)(x) ^ 0, J (Tf)(x) dx =
f f(x) dx. Tt admits infinitesimal generator A :
Ttf = exp (tA)f = strong lim^«, exp (nt((I — n~M.) -1 — /))/, — oo < t < oo.
Since (/ — n _ 1 A) _ 1 and (/ + n _ 1 A) _ 1 must be transition operators for n > 0,
(I — n~ A )~ exists as a transition operator. Hence A is the infinitesimal gen-
erator of a one-parameter semi-group St = exp (tA2), t ^ 0, of transition opera-
tors. Thus the Fokker-Planck equation in a Riemannian space R: df(t, x)/dt =
A2f(t, x), f(0, x) = f(x) G Li(R), t ^ 0, is integrable stochastically if A =
p%(x)d/dx% is the infinitesimal transformation of a one-parameter Lie group of
equi-measure transformations of R.
Extensions, i) If R admits several flows exp (tAt), i = 1, 2, • • • , m, with
mutually commutative Ai, then exp (t _£^ii2) defines a stochastic process in R.
ii) Let the group of motions in a Riemannian space R be a semi-simple Lie group
with infinitesimal transformations Xi, X2, • • • , Xm transitive in R. Then the
Casimir operator C = gtJXiX3 ((g%3) = (gi3)~l, g a = cipc%, [Xi, X3] = ci3Xk) is
commutative with every Xi. In this case, at least when R is compact, exp (tC)
defines a temporally and spatially homogeneous, continuous stochastic process
in R-^Sb Brownian motion in the homogeneous space R.

NAGOYA UNIVERSITY,
NAGOYA, JAPAN.