JOURNAL OF MATHEMATICAL ANALYSIS AND APPLICATIONS 12, f-12(1965)

Integrals of Set-Valued Functions*

ROBERT J. AUMANN

The Hebrew Unwersity, j%rusalem, Israel

Submitted by Samuel Karlin

INTRODUCTION

Set-valued functions have been af interest for some time. Fixed-paint

theorems for such functions were proved by Kakutani [I], Eilenberg and
Montgomery [z], and others; furthermore, set-valued functions have been
used repeatedly in Economics (see for example Arrow and Debreu [3],
McKenzie [43 and Vind [S]). Integrals of set-valued functions have been
studied in connection with statistical problems; see Kudo [6] and Richter [7].
Lately integrals of set-valued functions have arisen in connection with
economic problems [8, 91, and we here extend the basic theory of such
integrals.
Let T be the unit interval [0, 11. For each t in T, let F(t) be a nonempty
subset of euclidean n-space Es. Let +F be the set of all point-valued functionsf
from T to En such that f is integrable over T and f(t) EF(~) for all t in T.
Define
@t)dt = jS,f(t)dt:fe-1
i.e., the set of all integrals of members of 9. This notion is a natural generali-
zation of the integral of point-valued functions on the one hand, and of the
sum of a finite number of sets on the other hand. It is closely connected with
Vind’s “set valued measures” [S]; if for measurable subsets 5’ of T, we set
v(S) = J,F(t) dt, then v is a set-valued measure,
The following conventions will be used: Instead of STF(t) dt, j,f(t) dt,
etc., we will write SF, Jf, etc.; when it is necessary to integrate over a
s&et S of T we will write s,F, etc., but otherwise the range of integration
will not be specified and will be understood to be all of T. When we refer to
“all” or “each” t in T we will mean “almost all,” and when we say that
“there is a t in T” with some property, then we shall mean that the set of

* Research partially supported by the Office of Naval Research under contract

nun&x N62558-3586.

409 .I21 I-1

2 AUMANN

points t in T with the property in question is of positive Lebesgue measure.

“Measurable” will mean “Lebesgue measurable”. Coordinates of points in En
will be denoted by superscripts. If x and y are in En, x > y will mean1 xi 3 yi
for all i. The scalar product is denoted x * y, and 1x / means (1 x1 I, *a*, ) X%I).
Set-theoretic subtraction will be denoted \.
The following basic theorem is due to Richter [7]:

THEOREM 1. 1 F is con.vex.
It is natural to ask under what conditions SF (or equivalently, F) is non-
empty. The function F will be called Borel-measurable if its graph
((4 4 : x ~F(t)l is a Bore1 subse@ of T x En. It will be called integrably
bounded if there is a point-valued integrable function h from T to E” such that
1x 1 < h(t) for all .r and t such that x EF(~).

THEOREM 2. If F is Borei-measurable and integrably bounded, then SF is

nonempty.
Neither hypothesis in this theorem can be omitted, as we shall show by
counter-example. This theorem follows immediately from a basic “measurable
choice theorem” due to von Neumann [l l].
The function F will be called nonnegative if x > 0 for all x and t such that
x E F(t). For each t in T, F*(t) will denote the convex hull of F(t).

THEOREM 3. If F is nonnegative and Borel-measurable, then SF = SF*.

This theorem looks like a trivial consequence of Theorem 1, but it isn’t.
Neither hypothesis can be removed. Note that the theorem remains true if
F is bounded from below (or from above) by an integrable point-valued
function h, so that in effect the nonnegativity condition is a weakening of
the condition of integrable boundedness.
The function F will be called closed if F(t) is closed for each t.

THEOREM 4. If F is closed and integrably bounded, then SF is compact.

Note that the hypothesis of Borel-measurability is not needed here.
Theorem 4 was proved by Kudo [6] for F that are convex-valued and Bore1
measurable (in addition to the conditions given here). Richter [7] proved a
version slightly different from Kudo’s, which does not require convexity,
but which still requires measurability.

1 Note that this differs from the standard usage, in which this relation is denoted L.
la This definition is justified by the fact that a point-valued function is Borel-
measurable if and only if its graph is a Bore1 set (see [lo], p. 365, Proposition 4, and
p. 398, Proposition 2). For compact convex valued functions, this definition coincides
with that of Kudo [6] and Richter [7], who use the measurability of the support
function.
 INTEGRALS OF SET-VALUED FUNCTIONS 3

We now turn to a generalization of Lebesgue’s dominated convergence

theorem; this may be considered our chief result. If A, , A,, .*. are subsets
of En, then by definition [lo, p- 241 ff.], x E lim inf A, if and only if every
neighborhood of x intersects all the A, with sufficiently high k, and x E lim
sup A, if and only if every neighborhood of x intersects infinitely many Al,. If
lim inf A, = lim sup A, = A, then we write A = lim A, , or A, + A.

THEOREM 5. Zf Fk(t) + F(t) fov all t, and all the Fk are Borel-measurable
and bounded by the same integrable point-valued function, then J-F, ---f SF.
Theorem 5 may be restated in terms of continuous set-valued functions.
This will be done in Section 5. Neither of the conditions can be omitted in
this theorem. The proof of Theorem 5 makes use of two analogues of Fatou’s
lemma (Propositions 4.1 and 5.1), which are of some interest in their own
right.
In the final section, we consider an application to extreme points of sets
of vector functions, as treated by Karlin in [12].
Nowhere in this paper do we use the fact that the measure space T is
totally finite. In particular, the theorems remain true when T is (0, 00) or
(- 03, co). What is essential (to at least some of the theorems) is that T be
nonatomic.

PROOFOFTHEOREM 2

It will be convenient in this section to replace the euclidean space

En by an arbitrary separable and complete metric space X. The function
F will still be defined on T = [0, 11, but its values will now be subsets
of X. A point-valued function f from T to X will be called Lebesgue-
measurable if f-l(U) is a Lebesgue-measurable subset of T for every open
(or equivalently, Borel) subset of X. Recall that an analytic subset of X is
the continuous image of a Bore1 subset of X [lo, p. 3601; the set-valued
function F will be called a&~&c if its graph is an analytic subset of T x X.
The following proposition follows from a lemma of von Neumann [ 11, p. 448,
lemma 51:

2.1. If F is an analytic set-valued function from T to X, then

PROPOSITION
there is a Lebesgue-measurable point-valued function f : T -+ X such that
f(t) EF(t) for all t.
Since every Borel-measurable F is analytic, it follows by setting X = En
that if F is a Borel-measurable function from T to E”, then there is an f
satisfying the conclusions of Proposition 2.1. If, moreover, F is integrably
bounded, then it follows that f is integrable. Then jf E SF, and Theorem 2
is proved.
4 AUMANN

PROOF OF THEOREM 3

The proof is by induction on the dimension 71of the space. For II = 0 the
theorem is immediate, since F*(t) = F(t) = (0) for all t. Suppose the
theorem true for dimensions less than n. If the theorem is false in dimension
n, then for appropriate F we have SF* \ SF # 4; let x E SF* \ SF. By
Theorem 1, SF is convex, so it has a supporting hyperplane passing through
X; that is, there is a vector a E En such that

a-y<aax (3.1)
for all y E SF.
Since x E SF*, it follows that x = Jf*, where f*(t) eF*(t) for all t.
Furthermore, f* may be chosen Borel-measurable; for, every Lebesgue
measurable function is equivalent to (i.e., differs on a set of measure 0 from)
a Bore1 measurable function,2 and according to our convention, “all” means
“almost all.” Now recall Caratheodory’s theorem, which states that if D and
D* are subsets of En such that D* is the convex hull of D, then every point
of D* is a convex combination of n + 1 points of D (see Eggleston [13],
p. 34ff.). According to this, for each t there are positive real numbers
vo(th *.., vvL(t) summing to 1, and members g,,(t), .*., g%(t) of F(t), such that

f*(t) = &At) gdt). (3.2)

j=O

Wewish to show that the ‘pj and theg, obeying (3.2) can be chosen measurable,
and with go integrable. To this end, note that (3.2) implies that for at least
one of the gj , J$,g:(t) < EFclf*i(t). S’mce the indexing of the gi is of no
significance, this means that the subset G(t) of E(n+l)+n(n+l) defined by

G(r) = [(Q, ..‘, t,, x0, ‘.., x,) : 0 < Ej < 1 andxj EF’(t) for allj,

$Sj = 1, 3~: < $f*"(t), andf*(t) = $ [iXj/

i--o i=l i=l 3=0

is nonempty for all t. Furthermore, the graph of G is a Bore1 subset of

E(n+l)+~(pl+l)fl; this follows from the Bore1 measurability of F and f’. So
by Proposition 2.1, we may choose measurable yj and gi so that

(P)o@)9***t dt), go(t), S.-Ygrdt>> E G(t) for all t.

2 This is immediate for simple functions, and every measurable function is the limit
of a sequence of simple functions.
 INTEGRALS OF SET-VALUED FUNCTIONS 5

Then the yi and the gj are measurable and obey (3.2). Furthermore, from
g*(t) EF(~) and the nonnegativity of F it follows that gO{t) >, 0. Hence

and therefore the integrability of gi follows from that off*. Similarly all the
gi are integrable, i.e., g, is integrable.
We now show that
a *g&> < 62*f”(t) W)
for all t and j. Indeed, suppose that

Q - g?&) > a *f*(t) (3.4)

for some k and t, say for I E 5, where S has positive measure. For each t, there
is a j obeying
a * gJt) 2 a *f*(t); (3.5)
for otherwise, since vj(t) > 0, we have

“$1
U . f *(t) = $pj(t)
Q ‘giCt) < & Vj(t) u hf*(t) = a ‘f*(t),
i=l i=l

an absurdity. Let us denote the firstj that fulfills (3.5) byj(t). Define a func-
tion f by
when 2ES
when t $ s.

Clearly f is measurable, but possibly it is not integrable. For each positive

integer m, let U(m) = {t :f(t) < (m, *a*, m)}, and define a sequence of inte-
grable functions fm by

when t E U(m)

when t 4 U(m).

Then fm(t) E F(t) for all t, and so (3.1) yields

for allf. Now uz=, U(m) = T; therefore

s*\ti(n)Qd(80-f*) 4 0.
6 AUMANN

Furthermore, for sufficiently large m, U(m) 17 S has positive measure. For

such m, then, it follows from (3.4) that

In other words, if

then e(m) is monotone increasing in m, and c(m) > 0 for sufficiently large m,
say for m 3 m,, . Now from (3.7) it follows that for m sufficiently large, say
m > m, , we have

j a (fin -f*) = j r,V(m)

a*(go
-I*>z - 24mo)
Furthermore, by (3.5) and the definition off,

sU(m)\Sa.fm_ 1U(m)\SQ.f> $ U(m)\S

a -f*.

So if m > max (ml , m,), then

a.jfm- a.jf* =j a.(fflG

-f*)=ju~,~,s+
jvc,,,, -
i r\ U(m)

4%~ > 0
3 c(m,) i- 0 - 2 ,

contradicting (3.6). This demonstrates (3.3).

Next, we demonstrate
a *gj(t) = a *f*(t) (3.8)
for all t and j. Indeed, suppose that there is a j-say j = k-such that for
some t, strong inequality holds in (3.3) i.e., a . g9(t) < a *f*(t). Applying
(3.3) and vk(t) > 0, we deduce

a*f*(t)
=j$vi(t)
a <i=o
2 FiIt)
a-f*(t)
=a.f*(t),
'gjCt)

an absurdity. This proves (3.8).

Let H be the hyperplane {y : a . y = O}. Define E(t) = [F(t) -f*(t)] n H,
and let E*(t) be the convex hull of E(t). From (3.8) it follows that
&P> -f*(t) Ew and so by (3.2), 0 =f*(t) --f*(t) E E*(t). Since H is of
dimension n - 1, we may apply the induction hypothesis to E, and deduce
 INTEGRALS OF SET-VALUED FUNCTIONS 7

0 E s E. Let e be such that e(t) E E(t) for all t and se = 0. Then for each t,
e(t) + f*(t) E F(t), and J [e +f*] = Jf* = x. Hence x E JF after all,
contradicting our assumption. This completes the proof of Theorem 3.
To see that Theorem 3 is false if it is not assumed that F is Borel-measu-
rable, let it = 1, and let g be the characteristic function of a subset of T with
inner measure 0 and outer measure 1. Let F(t) contain the two points g(t)
and 2 only. Then SF = (2) but SF* = [I, 21.
To see that Theorem 3 is false without the nonnegativity assumption,
let n = 1 and let F(t) = (l/t, - l/t). Then SF = 4 and SF* = (- co, CO).

PROOFOF THEO= 4

The proof of Theorem 4 is a consequence of the following analogue of

Fatou’s lemma.

PROPOSITION 4.1. IfF1, F,, ..* is a sequenceof set-valued functions that

are all bounded by the sameintegrable point-valued function h, then
lim sup Fk 3 lim sup Fk .
s i
PROOF. Suppose x E lim sup SF,. Then x is a limit point of a sequence
sfk , wheref&j EFAt ) f or each K and t; that is, there is a subsequence of
sfk converging to x. We wish to show that x E J lim sup Fe; for this purpose
we may assume without loss of generality that x is actually the limit of the
s fk , i.e. that the subsequence converging to x is the whole original sequence.
The fk can be considered real-valued functions on (1, -me,EZ>x T; since
they are integrable, it follows that they are members of the Banach space
Ll =L’({l, .I*, n} X 7’). Because the fk are all bounded by the integrable
function h, it foilows that there is a subsequence with a weak limit, which we
call f (see Dunford-Schwartz [14], Theorem IV.8.9, p. 292). Again, we may
assume without loss of generality that fk actually converges to f weakly.
Let (9 , e2 , ..a} be a sequence of positive real numbers tending to 0.
Since {fi, f2, mm*}approaches f weakly, it follows that also {fm, f,,,+l, *.a}
approaches f weakly, for all m. Hence from a known theorem, it follows that
there is a sequence of convex combinations of fm , fm+l , *.. that approaches f
in the norm ofL1 ([14] Corollary V. 3.14, p. 422). For each m, letg, be such
a convex combination with I] g, -f 11< l m ) where 1111denotes the L1 norm.
Then g, -+ f in the norm of L1. Hence by a known theorem, there is a
subsequence of the g, that converges to f almost everywhere (see [14],
Theorem III, 3.6(i), p. 122, for a proof that norm convergence implies
convergence in measure, and Corollary III, 6.13(a), p. 150, for a proof that
convergence in measure implies convergence a.e. of a subsequence); without
8 AUMANN

loss of generality3 let it be the whole original sequence. So for all t,g,(t) +f(t),
and g&t) is a convex combination of {fm(t),fm+i(t), **a>.Since the latter are
points of En, it follows from Caratheodory’s theorem [13, pp. 34 ff.] that

where the tJi(t) are nonnegative and sum to 1, and e,,(t), a**,e,,(t) are chosen
from amongf,(t), fm+&>, -*- . Now for each t, we can choose a subsequence
of the g,Jt) such that all the corresponding subsequences of {f?,,(~)}, a+.,
PnmW>, k&t)>, *-a3and {e%,Jt)} converge. The limits of the 8’s in these
subsequences must be nonnegative numbers summing to 1, and the limits
of the e’s must be limit points of thef,(t). Hence for each t,

f(t) = lim gnz(t) = 2 e,e](t),

j=O

where the 0, are nonnegative and sum to 1, and the ej(t) are limit points of
{f1(t),f&>, -->- s o I‘f we let G(t) be the set of limit points of {fk(t)>, and G*(t)
the convex hull of G(t), then we have shown that f(t) E G*(t) for each t.
Hence JOE J G*.
As we have noted above, every Lebesgue-measurable function is equivalent
to a Borel-measurable function; so we may assume that the fk are Borel-
measurable. Then it follows easily that G is Borel-measurable. So we may
apply Theorem 3 and deduce that s G* = 1 G. Hence sf E s G. But since
fk(f) E Fk( t ) , 1‘t f o11ows that every limit point of the Fk(t) will be a member
of lim sup Fk(t), i.e., G(t) C lim sup Fk(t). Hence sf E 1 lim sup Fk . But
from the weak convergence of fk to f it follows that sf = lim jfk = x, so
that x E j lim sup Fk . This completes the proof of Proposition 4.1.
Suppose now that F is closed and integrably bounded; set Fl =
F, = . . . = F, Then lim sup Fk = cl F = F and lim sup SF, = cl SF, where
cl denotes “closure” (see [lo], p. 243, IV.6). So by Proposition 4.1,

and hence SF is closed. Since it is bounded by the integral of the function

h that bounds F, it follows that it is compact, proving Theorem 4.
It is possible to prove Theorem 4 somewhat more simply by a direct
application of [14, V.3.141; but Proposition 4.1 is interesting for its own sake,
and is needed in the proof of Theorem 5.

’ This involves cutting down the original sequence of fk’s and using the correspond-
ing subsequence of the gk’s.
 INTEGRALS OF SET-VALUED FUNCTIONS 9

If we replace the integrable boundedness condition in Theorem 4 by the

assumption that F is nonnegative, then SF need not even be closed. For a
counterexample, let 71= 2, let g(t) = ((1 - t)/t, t/(1 - t)), and let
F(t) = (O,g(t)}. Then F is the union of the open positive quadrant with the
origin (0).

PROOF OF THEOREM 5

We first prove another analogue of Fatou’s lemma, as follows:

PROPOSITION 5.1. If all the Fk are Borel-measurable and bounded by the

same integrable point-valued function, then

lim inf Fk C lim inf Fk .

s s
PROOF. Let x E l lim inf Fk . Then x = Jf, where f(t) E lim inf Fk(t)
for each t. Since f is equivalent to a Borel-measurable function, it may be
assumed to be Borel-measurable. Now the space En x E” x +.* may be
metrized so that it is complete and so that its topology is the usual product
topology [IO, p. 3131. For each t, define a subset G(t) of En x E” x *** by

G(t) = ((~1, $2, -1 : x1 ~Fi(t), x2 eF,(t), .--, and lim xk =f(t)>.

Then the statement f(t) E lim inf Fk(t) is precisely equivalent to the statement
G(t) +b ~10, P. 2421, and hence G(t) #d, for each t. Furthermore, G is
easily seen to be Borel-measurable and hence analytic. So by Proposition 2.1,
there is a measurable functiong from T to En x En x .a. such thatg(t) E G(t)
for each t; that is, a sequence fi , fi , a.. of measurable functions from T to
En, such that fk(t) E Fk(t) for each t, and limfk(t) = f(t). Now since
fk(t) E F,(t), all the fk are bounded by the same integrable point valued
function. Hence from Lebesgue’s bounded convergence theorem, it follows
that lfk + sf = X. But sfk E SF*, and so x t lim inf JFk [lo, p. 2421.
This completes the proof of Proposition 5.1.
If F(t) = lim infF,(t) = lim sup Fk(t) for all t, then by Propositions 4.1
and 5.1,

s I
F = lim infF, C lim inf
i
Fk C lim sup lim supF, =
j
F;

hence equality holds throughout, and so lim sFk exists and equals slim Fk .
This proves Theorem 5.
Theorem 5 is false without the Borel-measurability assumption. Indeed, let
n = 1, let g be the characteristic function of a subset of T with inner measure
 10 AUMANN

0 and outer measure 1, and let Fk(t) = {g(t)/k}. Then lim SF* = 4 but
S lim F, = (0).
Let A be an arbitrary subset of a metric space X, and let G, be a set-
valued function defined for x in A, whose values are subsets of En. G is called
upper-semicontinuous if x, + x implies G, r> lim sup G, for all x, and x
in A; lower-semicontinuous if x,+x implies G, C 1im”inf Gz, for all x,,
and x in A; and continuous if x, ---f x implies G, = lim G, for all x, and x
in A. This is the same as the standard definition of the& terms (see, for
example, Karlin [15], p. 409).

COROLLARY 5.2. Let F,(t) be a set-valued function defined for t E T and

x E A, all of whose values are bounded by the sameintegrable point-valued func-
tion, and such that F, is Bore&measurable for each Jixed x E A. Then sf F,(t)
is upper-semicontinuous in x for eachfixed t, then J F, is upper-semicontinuous;
zf Fz(t) is lower-semicontinuous in x for each fixed t, then SF, is lower semi-
continuous; and if F%(t) is continuous for each fixed t, then SF, is continuous.

PROOF. Follows from Propositions 4.1 and 5.1.

APPLICATIONS TO EXTREME POINTS OF SETS OF VECTOR FUNCTIONS

In this section, we make use of Proposition 2.1 only; the other results will
not be used. Let A be a compact convex subset of En, B the set of its extreme
points, cl (B) the closure of B. If n > 3, it is not necessarily true that
cl (B) = B. Let AA, A’, and AC1 (Bj be the sets of all measurable functions
from T to A, B, and cl (B) respectively. Since the space of all measurable
functions from T to En has a linear structure, we may discuss the extreme
points of k’A . In [12], Karlin proved that the set of extreme points of &A
includes Ma and is included in AC1 (Bj .

PROPOSITION 6.1. The set of extreme points of JZA is precisely AB .

PROOF. Clearly, every point of JZB is an extreme point of &A . Conver-
sely, let f be an extreme point of 4A . If f is not in &B , then for some t,
f(t) is not an extreme point of A. Hence for each t we may choose g(t) and
h(t) in A so that f(t) = *g(t) + *h(t), and g(t) and h(t) differ from f(t) for at
least some t. Because of Proposition 2.1, g and h may be chosen so as to be
measurable as well. Then g and h are in J?~ and f = *g + 8 h, contradicting
the extreme point property off. This proves the proposition.
Another situation treated by Karlin in [12] is the following: Let p1 , ..., pna
be a set of nonatomic totally finite measures on T, and let a, , a-., a,,
b1, *a., b, be in E”. Let A and B be as above. Let 3 be the subset of .MA
 INTEGRALS OF SET-VALUED FUNCTIONS 11

consisting of functions f such that 6, < sfdpi < ai for i = 1, **a, m. Karlin
proved that the extreme points of 9 are contained in AC1 (BJ.

PROPOSITION 6.2. The extreme points of 3 are contained in &ZB.

PROOF. The proof follows Karlin’s ideas, but use of Proposition 2.1 makes
it simpler and yields the stronger result. Let f be an extreme point of 3,
and suppose it is not in dB . Construct g and h as in the previous proof, and
lete=&g-*h,sothatf+e=gE A*andf-e=hh AA.ForSCT,
set CL(S) = {Js edpI , *a*, Js edp,]. Then p is a vector measure of dimension
nm. Applying Lyapunov’s theorem on vector measures [16], we obtain
a subset S of T such that p(S) = -8p(T). Define e’ on T by e’(t) = e(t)
for t E S, and e’(t) = - e(t) for t $ S. Now define fi and fi by f, = f + e',
fi = f - e’. Then f = Bfi + if2, fi and fi are in J&” , and

= If& + */edpi - ij edpi = 1 fdpi.

Hence fi E 9. Similarly f2 E 59, and the proof is complete.

If F is the set-valued function defined by F(t) = A for all t, then F = AA.
Propositions analogous to 6.1 and 6.2 can be proved in the general case, when
F need not be constant, but is assumed to have compact convex values. In
this connection, we note that Theorems 3 and 4 of this paper are generaliza-
tions of Theorem 3 and Corollary 1 of [12].

ACKNOWLEDGMENTS

We gratefully acknowledge some very helpful conversations with Lester Dubins,

Yakar Kannai, Bezalel Peleg, and Micha Per& Also, we wish to thank Gerard Debreu
for calling our attention to some of the references.

1. S. KAKUTANI, A generalization of Brouwer’s tied point theorem. Duke Math. J.

8 (1941), 457-459.
2. S. EILENBERG AND D. MONTGOMERY, Fixed point theorems for multi-valued trans-
formations. Am. J. 1Math. 8 (1946), 214-222.
3. K. J. ARROW AND G. DEBREU, Existence of an equilibrium for a competitive
economy, Econometrica 22 (1954), 265-290.
4. L. W. MCKJXNZIE, On the existence of general equilibrium for a competitive
market, Econometrica 27 (1959), 54-71.
5. K. VIND, Edgeworth allocations in an exchange economy with many traders.
Econ. Rev. 5 (1964), 165-177.
12 AUMANN

6. H. KUDO, Dependent experiments and sufficient statistics. Nat. Sci. Rept. Ocha-
nomizu Univ., Tokyo 4 (1954), 151-163.
7. H. RICHTER, Verallgemeinerung eines in der Statistik be&t&ten Satzes der
Masstheorie. Math. Annalen 150 (1963), 85-90 (for correction to this article see
same volume, pp. 440-441).
8. R. J. AUMANN, Existence of a competitive equilibrium in markets with a con-
tinuum of traders. Econometrica, in press.
9. R. J. AUMANN AND M. A. PERLES, A variational problem arising in economics.
J. Math. Anal. Appl., in press.
10. C. KURATOWSKI, “Topologie I.” Monografie Matematyczne, Warsaw, 1948.
11. J. VON NEUMANN, On rings of operators. Reduction theory. Ann. Math. 50 (1949),
401-485.
12. S. KARLIN, Extreme points of vector functions. Proc. Am. Math. Sot. 4 (1953),
603-610.
13. H. G. EGGLESTON, “Convexity.” Cambridge, Univ. Press, 1958.
14. N. DUNFORD AND J. T. SCHWARTZ, “Linear Operators, Part I.” Interscience,
New York, 1958.
15. S. KARLIN, “Mathematical Methods and Theory in Games, Programming, and
Economics,” Vol. I. Addison-Wesley, Reading, Mass., 1959.
16. A. LYAPUNOV, Sur les fonctions-vecteurs completement additives. Bull. Acad.
Sci. URSS, St. Math. 4 (1940), 465-478.