BON~XCE,W. and V. L. KLE~.

,
Math. Annalen 152, 1--29 (1963)

The Generation of Convex Hulls*

By
WmHAM BONNICEin Durham (New Hampshire) and
VICTOR L. KLEE in Seattle (Washington)

§ O. Introduction
When S is a subset of a (real) linear space E, the convex hull of S (conS)
is the intersection of all convex sets in E which contain S. Equivalently,
m
cons = U conmS, where x ~ conmS iff x = ~ :qs~ for some st E S and
l_~m<~ 1
m
ai ~ 0 with ~ :¢~ = 1. When E is a locally convex Hausdorff linear space, the
1
closed convex hull of S (cl conS) m a y be described alternatively as the inter-
section of all closed convex sets containing S, as the closure of the convex hull
of S, and as the set of all points p ~ E such t h a t / p <_- sup I S for all / C E*
(the space of all continuous linear functionals on E). The present study is
devoted to some other hull-forming operations which are closely related to con
and cl con. We are interested in the extent to which these operations resemble
con and cl con, and in inter-relationships among the various operations. Section
headings are as follows:
§ 1. The operations conm.
§ 2. A generalization of theorems of Carath~odory and Steinitz.
§ 3. Types of positive independence.
§ 4. Barycenters of measures : Setting and general results.
§ 5. Finitely additive measures.
§ 6. Countably additive measures.
Most of the results of §§ 1--3 are purely algebraic and hence valid in vector
spaces over arbitrary ordered fields. The operations H considered in §§ 1--2
are such that always S ( H S ( conS. The main result of § 1 is the remark that
always con~(con~S)= e o n ~ S , which in conjunction with Carath6odory's
theorem leads to an easy proof and a generalization of a well-known result
on the generation of convex hulls by the formation of segments. § 2 defines the
d-interior of a set in such a way that for each convex subset X of R ~, int0X = X
and intnX is the ordinary interior of X relative to R n. I t is proved that if
X ( R n and p C inta coaX, then p E intd con Yd for some set Ya C X with

* The work of both authors was supported in part by the National Science Foundation,
USA (NSF-G18975). The paper is based on Dr. Bo~Ie~?s 1962 Ph. D. thesis at the
University of Washington, written in collaboration with Prof. KLE~,.
Math. Ann. 152 I
2 WILLIAM BONNICE a n d VICTOR KLEE :

cardY~ ~ max(2d, n ÷ 1). This result appears to be new for O < d < n ,
though for d = 0 and d = n it is equivalent to well-known theorems of CA~A-
TH~ODORY [11] and STEI~TZ [38] respectively. § 3 treats the relationships
among various types of positive independence and in particular it solves a
problem recently proposed by McK~N~Y [31]. I t is related in spirit to much
of § 2 but makes no other contact with the rest of the paper.
The last three sections are devoted to the notion of the barycenter of a
measure, this being a generalization of the notion of convex combination.
Emphasis is laid upon certain hull-formations which, in contrast to those of
§§ 1--2, include more than the ordinary convex hull. Since the notions of these
sections are more technical than those of §§ 1--3, the reader is referred to § 4
for further introductory material and to §§ 5 - - 6 for the theorems themselves.

§ 1. The operations eon,~

For a set X in a linear space and for each positive integer m let conmX
denote the set of all convex combinations of m or fewer elements of X, i.e.,

conmX = eix~ : ~i > O, ~ ' ei = 1, x~ C X .

1

Equivalently, eonmX is the union of all j-simplices which have their vertices
in X for 0 ~ j ~ m - 1. Thus eonX = U eonmX, and Carath~odory's theo-
1
rem [11] asserts that for X C R n, c o n X = c o n . + l X . The question arises, for
each fixed integer m with 2 ~ m ~ n ÷ 1, how m a n y times must the operation
eonm be iterated to produce the convex hull of a set in R n ? The case m -- n Jr 1
is settled by Carath~odory's theorem. For m = 2 the question asks, in order to
form the convex hull of a set X ( R ~ how m a n y times is it necessary to repeat
the operation of forming all line segments which have end points in X, then
forming all line segments which have end points in the resulting set, etc. ? This
ease has been treated b y several authors (cf. p. 10 of BONNES~-FE~c~EL [8])
and most recently by ABE-KuBoTA-Yo~EGUCHI [1]. The following general
result includes the answer to the above question as a special case. I t is an
immediate consequence of Carath~odory's theorem together with the observa-
tion that always con~(eonkX) = eon~ ~X.
1.1. Theorem. For X C R ~ i / ] ~ j ~ _ ~ . . . ]1 ~ n -~ 1 then
(co,j,X)...)) = conX.
I n order to prove 1.1, we first establish
i
1.2. Lemma. I n any linear space let X = U X r where each Xr is nonempty.
1
I] a E c o n X then there exist ar E conXr and 2r ~ 0 with ~ 2r = 1 such that
1
J
a = X It, at.
1
 The Generation of Convex Hulls 3

Proo/. If a is an element of con X then a is a convex combination of

finitely m a n y elements of X, so a can be written
Sl S8 S/

a=~ixi+ ~ ~ixl+'''+ ~ ~x,

1 sl+l sl-x+ i
8f
where each x~ ~ 0, ~ ~, = 1, and
1
{xl, x~. . . . . x,,} : X i ,

{x,,+l, x8,+2 . . . . . xs,} < X2,

• . . , , . , , . . . . . . . . . .

{Xsj_,+ 1, % , + ' 2 ..... XSj} < X j .

8r
N o w for r = 1, 2 . . . . . j let ~, = ~ ~i (where s o = 0). T h e n each 2~ ~_ 0 a n d
st-l+ 1
y
~r = 1. F o r those r such t h a t ~r > 0, let
1

2
r 8r-l÷l
con .
i
F o r those 2r such t h a t 2r = 0, let a r be a n y m e m b e r of c o n X r, Then a = ~ 2,a~,
1
a representation for a of the desired form. []
W e can now prove the basic lemma, thus establishing Theorem 1.1.
1.3. Lemma. For X in any linear space and positive integers ~ and k,
eonj (con~X) = eonj kX.
J i
Proo]. I f a { coni(conkX ) t h e n a = ~ ~rYr where ~r ~ 0, ~ 2r = 1, and
1 1
k k
Y~E eonkX, say y ~ = ~ fl~,xr, with { ~ , ~ 0, ~ f l ~ , - - 1 , and Xr, E X. T h e n
s=l s=l
j k
a= ~ ~ 2rflr,a:r,, a convex combination of ~k elements xr, of X, so
r=ls~l
a E conj~X.
jk
On the other hand, if a E c o n j k X t h e n a = ~ a,x, with each at ~ 0,
1
jk
ai = 1, and each x, C X. Break the set {Xl, x 2. . . . . x~.k} into ~ sets Xr with
1
each Xr having k elements. B y 1.2 there exist ar ~ conXe = con, X r and
2r => 0 with ~ 2, = 1 such t h a t a = ~ ,l~a,. Thus a { eon~(eon, X).
1 1
(Note t h a t the p r o p e r t y of convex combinations expressed b y 1.3 is possessed
also b y l~ositive combinations a n d b y linear combinations. However, the be-
havior of alfine combinations is more complicated (KLm~ [28]).)
N o w with X C R ~ a n d ~ _ i . • • Ji ~- n + 1, we have

eon~r(conh_,(... ( c o n h X ) . . . ) ) = cOnirir_,...hX) c o r m + i X = X ,
1"
4 WILLIAM BONNICE and VIc~rORKLEE :

where the first equality is justified by 1.3 and the second by Carath~odory's
theorem. Thus 1.1 is established, and a precise answer can be given to the
question posed at the beginning of this section: How m a n y times must the
operation con,n be iterated to produce the convex hull of a set in R" ? For any
set X in a linear space let con~X = conmX and for integers r > 1 let con~nX
-- conm (con~-1X). From 1.3 it follows that con~r~X = conm,X and as a special
case of 1.1 we can state
1.4. Corollary. I / X C R '~ and m ~ ~ n ÷ 1 then conVeX = conX.
Theorem 1.1 and its corollary 1.4 are best in the sense that the number
n ÷ 1 cannot be reduced for arbitrary sets X C R% This can be seen b y t a k i n g X
to be the set of n Jr 1 vertices of an n-simplex in R n. However if certain con-
ditions are imposed on X then n ÷ 1 can be reduced to n in 1.1 and 1.4. For
connected X this is mentioned b y BONNESE~-FE~cHEL [8]. A more general
reduction depends on theorems of FE~CHEL, BtTNT, and HA~NnR-RhDST~5~
[22], summarized in [22] as follows:
1.5. Theorem, I / a se~ X i n R n has at most n components ( i n particular,
i / X is connected) or i/ X is compact and has at most n convexly connected com-
ponents then every point i n the convex hull o / X is a eoT~vex combination o / n or
/ewer elements o / X .
(A set Y is called convexly connected if there is no hyperplane H such that
H ~ Y = 0 and Y has at least one point in each open halfspace determined
by H. Every connected set is convexly connected but not conversely. A maximal
convexly connected subset of a set Z is called a convexly connected component
of Z.)
From 1.5 in conjunction with 1.3 we obtain
1.6. Theorem. 1 / X C R ~ is as in 1.5 and i/?,.Jr-1 • • • Jl ~ n then
coni, (coni,_, ( . . . ( c o n j ~ i ) . . . ) ) = con X .
1.7. Corollary. I / X C R n is as in 1.5 and i / m r ~ n then c o g ~ X = conX.
Even for compact connected X, it m a y be impossible to reduce the number n
in 1.6 and 1.7. For example, let S be an n-simplex in R n with one vertex at the
origin 0 and let X be the union of those edges of S which emanate from 0.
Then X is compact, connected, and c o n X == S, but no point of the interior of S
is a convex combination of fewer than n elements of X.

§ 2. A generalization of theorems of Carath~odory and Steinitz

Let us recall the following welt-known theorems, due respectively to C)mA-
TH~ODORY [11] and ST~I~ITZ [38] (see DA~ZER-GRi)NBAuM-KL~ [14] for
alternate proofs and for references to other proofs in the literature):
I / X C R n and P E conX, then p E con Y /or some Y C X with card Y
~n+l.
I ] X ( R '~ and p E i n t c o n X , then p E i n t con Y /or some Y c X with
card Y ~ 2 n.
This section contains a common generalization of the two results, based on
the notion of the d.interior of a set X (in a linear space), where int~X is the set
 The Generation of Convex Hulls 5

of all points p such that p is in the relative interior of some d-simplex contained
in X ; equivalently, p E int~X iff there exists a d-dimensional flat F such that p
is interior to X r-~ F relative to F.
The following remarks are obvious or follow from familiar arguments:
2.1. Proposition. Always int~(int~X)= intmax~d,e)X, SO the operator int~
is idempotent. For X C R n,
X = int oX ) int 1X ) . • • ) int n_ 1X ) int~ X = int X ) 0 = intn + 1X .
2.2. Proposition. I / the convex set C is supported by the hyperplane H, then
int d (C f~ H) = (int~ C) f~ H.
2.3. Proposition. I / P is a convex polyhedron, inta P is the disjoint union
o/all/aces o/ P o/dimension ~ d (where these/aces are open relative to their own
a/fine hulls).
2.4. Proposition. I / 0 ~ d ~ n + 1 and M is the set o/ all vertices o/ an
n-simplex conM, then condM and intdconM are disjoint sets whose union is
conM.
The principal result of this section, first stated by BONNICE [9], is
2.5. Theorem. I / X ( R n, 0 <~ d -< n, and p C intaconX, then p ~ intacon Ya
/or some Yd C X with card Ya g max (n + 1, 2d).
We see from 2.1 that this becomes Carath6odory's theorem when d = 0 and
Steinitz's theorem when d = n. Note that if p is interior to an n-simplex c o n X
with vertex-set X, then Y~ must contain all n + 1 points of X, while if B is a
linear basis for R n, X .= B ~ ( - B), and p = 0, then 2d points are needed for
Yd. Thus 2.5 cannot be improved for general X ( R n.
The proof of 2.5 can most easily be effected in terms of the following
homogeneous formulation, where p o s X is the set of all points of the form
k
o,ix i with ai --~ 0 and xl E X . The fact that 0 E intdcon Y iff pos Y contains
1
a linear subspace of dimension d is a connecting link between 2.5 and
2.6. Theorem..For pos X = R ~ and 0 <-- d <---n, there exists Yd C X with
card Y~ g m a x ( n + 1, 2d) such that pos Ya is a linear subspace o~ R n o/ di-
mension >=d.
I n the discussion below, we first derive 2.5 from 2.6, then prove 2.6, and
finally describe some other applications of 2.6.

Suppose X C R ~, 0 <- d -<- n, and 0 ~ int~conX. Then 0 lies in the relative

interior of a d-simplex which is contained in conX, and since each of its
vertices is a convex combination of some finite subset of X, we conclude
that 0 E intaconX' for some finite X ' C X . Let X o be a set of smallest
cardinality for which 0 E int, conXo. Let E be the linear hull of X 0. From 2.2
and the minimality of X 0 it follows that X 0 is not bounded by any hyperplane
through 0 and consequently (using the support theorem for convex cones)
p o s X 0 = E. Applying 2.6 with the roles of its R" and X played by E and X o
respectively, we obtain Y, C X 0 such that card Y, < m a x ( n + 1, 2d) and
6 WILLI~ Bol~Ic~ and VICTORKLEE :

pos Y~ is a linear subspace of dimension e ~ d. B u t then

0 E inte con Y~ C int~ con Y~
and it follows t h a t 2.6 implies 2.5.

The proof of 2.6 depends on two lemmas, of which the first is

2.7. L e m m a . Suppose L and S are supplementary linear subspaces o/ the
linear space E, and ~ is the linear projection o/ E onto S whose kernel is L. Then
/or all Z C E, L + p o s g Z = L ÷ posZ. Thus i / L = p o s M , then
p o s ( M ~J z Z ) = p o s M + p o s ~ Z = p o s M ÷ p o s Z = p o s ( M ~J Z ) .
k
Proo/. Consider first a point p ~ L ÷ p o s z Z , p having the form x ÷ ~ fl~gzi
1
with x E 35, fl~ ~ O, and z i E Z. E a c h zi has a unique expression in the form
xi ÷ z z i with xi E L, and thus we have

p=x+2j~t,~z~= x- ~ix~ + ~i~L+posZ.

1
k
Similarly, if q = x + ~ ~ z i E L + posZ, then with z i . = xi + ~zi we have
1

q= x+~YTe~x~ +~YTe~z~EL+pos~Z.n
1 / 1

W h e n L a n d 3 / a r e subsets of a linear space, we will say t h a t L is mini-

maUy spanned b y 3'/iff L is a finite dimensional linear subspaee, pos M = L, and
c a r d 3 / = (dimL) + 1.
2.8. L e m m a . Suppose M and Z are subsets o / a linear space satis/ying the
/oUowing conditions :
pos M is a linear subspace o/dimension e;
Z is the vertex.set o/ an /.simplex c o n Z ;
p o s M • e o n Z = {p} C i n t l c o n Z .
Then
(i) pos M f~ affZ = (p} (where aft Z is the smallest fiat containing Z) ;
(ii) pos (M ~J Z) is a linear subspace o/ dimension e ÷ /;
(iii) i / p :4:0 and i / p o s M is minimally spanned by M, there exists a set N
such that Z C N C M ~J Z, p o s N is minimally spanned by N, and dim (pos/V) >=
>/+1.
Proo/. The first assertion is left t o the reader. F o r the second, let S be a
subspace supplementary to pos M and let g denote the linear projection onto S
whose kernel is p o s M . T h e n the restriction of ~ t o affZ is biunique. Indeed,
if u, v E affZ with g u = gv, t h e n
~(u-v÷p)--~u-~v÷~p--O, whence u-v÷pEposM.
B u t also u - v ÷ p E affZ, whence b y (i), u - v ÷ p -~ p so u = v. Since the
transformation ga~z is a/fine and binnique, it preserves/-interiors, and since
 The Generation of Convex Hulls 7

P E intsconZ it follows that 0 C intlcon~rZ. Consequently posztZ is an /-di-

mensional linear subspace (minimally spanned b y z Z ) and (by 2.7)
pos(M ~ Z) = p o s M ÷ pos~rZ.
Since p o s M + p o s ~ Z is an (e ÷/)-dimensional linear subspace, {ii) has been
established.
For the third assertion, we note that since p o s M is a linear subspace, then
posM = p o s ( - M ) and the point p # 0 in p o s M admits an expression in the
fol?m

(~) p = 2: ~ ( - m ) ,
mEM

where all the ~ns are > 0 with at least one (but not all) = 0. Let V = {m: ~,~ > 0}
and N = Z ~ V. Then the origin 0 is a strictly positive combination of N.
Indeed, we note that since p E intf conZ, p has an expression in the form
(2) p = Z ~z
zEZ
with fl~ > 0 and ~ flz = 1. But then of course
zEZ

zEZ vEV
To complete the discussion of (iii), it will suffice to prove that when posM
is minimally spanned by M, the set N is a/finely independent.
Consider, then, numbers Xn for n ~ h r such that
(3) Z~z+ ~v=0
zEZ vEV
while
(4)
zEZ vEV

We wish to show that ),~ = 0 for all n, and will prove first that X = 0, where
4. = z~ ~- If 2 # 0 then, from (3),
zEZ
Z (~I~)~ = - Z (~vl~)v,
zEg vEV
a point of the set
(5) affZ ~ posM = {p)
(using (i) and the fact that posM is a linear subspace). Since Z is a/freely
independent and V is hnearty independent (by the assumed minimality of M),
it follows from (1) and (2) that always ~,/X = flz and 2v/~ = x~. But then

zEZ vEV z EV ]

in contradiction of (4). Consequently ~ = 0.

From (3) we have

zEZ vEV
8 W ~ L I ~ Bo~rm~. and VICTORKLEE:

where (with 2 = 0) the left-hand point is in affZ and the right-hand point is
in the linear subspace posM. But then from (5) it follows t h a t

X , ~ z = o = Z : ~vv.
zEg vEV
From the affine independence of Z and the linear independence of V we
conclude that 2n = 0 for all n E N. [~
Finally ready for the proof of 2.6, we restate the theorem for convenient
reference : F o r pos X = R • and 0 <- d ~ n, there exists Y g C X with card Y~
m a x ( n ÷ 1, 2d) such that pos Yg is a linear subspace o / R '~ o / d i m e n s i o n ~ d.
Proo[. The theorem is obvious in R 1. Assume it holds in R k for all k < n,
and consider the situation in R ~ as described with posX = R ~. For each
integer d E [0, n], we wish to produce a Yd with the desired properties. Assume
without loss of generality that 0 (~X and let W be a subset of X which has
minimum cardinality compatible with the requirement that 0 E con W. We
claim that the linear hull L of W is minimally spanned by W. First, pos W = L,
for otherwise W is bounded by a hyperplane H through 0, whence
0 E con (W ~ H) in contradiction to the minimality of W. To see that card W
= ( d i m L ) + 1, note that b y Carath~odory's theorem 0 E con V for some
V C W with cardV=< ( d i m L ) + l . I t follows that V = W and thus L is
minimally spanned b y W.
Now let M be a subset of X which has maximum cardinality compatible
with the requirement that p o s M is minimally spanned by M, and let e be the
dimension of pos M. Then M can serve as Ya for all d g e, and we may assume
henceforth that d > e + 1. Let S be a linear subspace of R n supplementary to
posM and let ~t be the projection of R n onto S whose kernel is posM. Then
S = pos~t(X ~ M) and by the inductive hypothesis as applied to S there
exists Za_ ~ C X ~ M such that
(6) cardZa_e < max(n - e ÷ 1, 2(d - e))
and pos~Z~_ e is a linear subspace of S of dimension / > d - e. Writing Z for
Z~_e, we see from 2.7 that
p o s M + p o s ~ Z = pos(M ~J Z ) ,
so pos (M ~J Z) is a linear subspace L of R ' , with dim L = e + / > d. I f e ÷ ] # n,
the inductive hypothesis as applied to L guarantees the existence of a subset ] 7
of M ~j Z with card Y d ~ max(n, 2d) such that pos Ya is a linear subspace
of L of dimension ~ d. Thus we m a y assume henceforth that ] = n - e and so
L = R". Since c a r d M = e + 1, the proof can be completed by establishing
(7) e + 1 + cardZ < m a x ( n + 1, 2d),
and since
e+l+2(d-e)=2d-e+l_~ 2d,
attention m a y be restricted to the case in which cardZ > 2 (d - e), whence
(by (6)) cardZ g n - e + 1. With
dim p o s z Z = [ = n - e,
 The Generation of Convex Hulls 9

it follows t h a t
cardZ = cardaZ = n - e + 1
(and pos~rZ is minimally spanned b y ~Z). Thus
e+l÷cardZ=n4-2
so inequality (7) reduces to
(8) n 4- 2 E m a x ( n ÷ I, 2 d ) .
To show t h a t 2.8 applies we have 0 = ~ a~zz with a~ > 0, ~ a~ = 1, and,
zEZ zEZ
since Z is also affinely independent, p . = ~ c¢~z is relatively interior to the
zCZ
/-simplex conZ. Also z p = 0 so p o s M f~ c o n Z ) {p}. F o r the reverse inclusion,
suppose q is a n y point of p o s M ;~ conZ, say q = ~ fizz with flz > 0 and
zEZ
flz = 1. F r o m the fact t h a t
zCZ

zEZ
and the affine independence of 7~Z, we see t h a t always fl~ =~ a z. Thus q = p and
p o s M r~ c o n Z = {p} C intfconZ.
If p = 0 then posZ is a linear subspace (of dimension / =: n - e) minimal[y
spanned b y Z and supplementary to p o s M . This and the choice of e show t h a t
e > n - e, and since d > e 4- 1, it follows t h a t
n-i-2<= n-~- 2 ( d - e ) ~ 2 d .
T h u s (8) is established if p = 0.
I n the case p ~ 0, 2.8 (ii) shows t h a t again R ~ = p o s ( M ~y Z). This time the
m a x i m a l i t y of e and 2.8 (fii) insure t h a t e ~ (n - e) 4- 1 whence n + 2 (d - e)
~2d-1.2

U n d e r certain circumstances the cardinality of Ya m a y be m a d e smaller

t h a n it was in 2.6. I n fact the proof of 2.6 goes t h r o u g h to prove just such a
result which we now state a n d t h e n use to prove a t h e o r e m in the literature.
2.9. Theorem. I / p o s X = R'* then there is a subset M o / X such that p o s M
is minimally spanned by M. Suppose M has maximum cardinality with respect
to this property and suppose e is the dimension o/ posM. T h e n / o r 0 < d < e
there is a subset Ya (= M) o/ X such that pos Ya is a linear subspace o/dimensione.
For e 4- 1 ~ d g n there is a Y~ C X such that pos Yd is a linear subspace o/
dimension ~ d and
cardYa~i2d-e4-1 < 2d
or
cardYa=n4-2~ 2d with e ->- n - e and p o s Y a = R " .
A result still stronger (and more complicated) t h a n 2.9 can be w r u n g from
the proof of 2.6. However, 2.9 suffices t o prove t h e sharper form of Steinitz's
Theorem first given b y STEINITZ himself [38] a n d discussed also b y C. V. RO-
BINSON [34]. T h o u g h our proof is longer t h a n t h a t of STI~I~ITZ, we include
it to illustrate the applicability of 2.9.
10 WILLIAM Bom~IC~, and VICTOR KL~E:

2.10. Corollary. I / p is interior to c o n X C R n and i / n o three points o / X

lie on any line through p (in particular, i / X is a subset o/the boundary o/ c o n X )
then either
(i) X consists o/exactly 2n points collinear in pairs with p, or
(ii) there is a Y ( X with card Y g 2 n - 1 such that p is interior to con Y.
Proo/. We assume w i t h o u t loss of generality t h a t p = 0 ~ X and then we
prove t h a t either (i) holds or there is a set M C X such t h a t pos M is a linear
subspace of dimension ~ 2 minimally spanned b y M. (This proves 2.10 for
n g 2 and also shows t h a t e >_- 2 in 2.9, whence, for n ~ 3, 2.10 follows from
2.9.)
I f there is an x 0 E X such t h a t the r a y ] - c~, 0 [ x 0 misses X then this r a y
intersects t h e b o u n d a r y of con X in some point q ~ X. There is a hyperplane H
t h r o u g h q supporting c o n X and q ( con (H ~ X). B y Carath~odory's theorem,
q is in t h e relative interior of some It-simplex with vertex set V ( H ~, X and
with 1 g k ~ n - 1. I n this case {x0} ~ V will do for our set M.
I f ] - 0 % 0 [ x intersects X for all x E X we m a y assume t h a t X = - X .
Since l i n X ) p o s X = R n there is a subset B of X which is a linear basis for R '~.
If X = B ~ ( - B) then situation (i) prevails. Otherwise let a E X ~ (B ~ ( - B)).
Since p o s ( B ~ ( - B)) = R n there exist ~b ->- 0 and fib ~ 0 such t h a t

= Z ( bb + b)).
bEB

If ~b ~ fib write ~b b ÷ fib ( - b) = (~b - fib) b, and if ~b < fib write ~b b ÷ fib ( - b)
= (fib - ab) ( - b ) . I n this w a y we obtain a positive representation for a in
which, for each b ~ B, at m o s t one of b a n d - b has a non-zero coefficient. Thus
there is a linearly independent set C C B ~ ( - B) with 2 g card C g n such
t h a t a = ~ 7cc with each 7c > 0. Then pos((a} ~J C) is a linear subspace of
eEC
dimension card C minimally spanned b y (a) w C. F o r a proof of this last fact
see, for example, G ~ ] ~ [19]. []
Two open problems seem w o r t h y of m e n t i o n :
(I) Find an analogue o / 2 . 1 0 in which int c o n X is replaced by int~ c o n X .
(II) Find a result which is related to 2.5 as 1.5 is related to Carathdodory's
theorem.
Examples indicate t h a t in (II), one should p r o b a b l y require p E (int c o n X ) ~
~ X r a t h e r t h a n merely p E int~conX. T h u s the problem takes the following
form. F o r 0 ~ d g n, find the smallest number/c = k (n, d) which has the follow-
ing p r o p e r t y : W h e n e v e r X is a connected set in R n and p E ( i n t c o n X ) ~ X,
t h e n p E int~con Y for some Y C X with card Y g h. B y 1.5, k(n, 0) = n. B y
t a k i n g X t o be the union of all t h e edges of an n-simplex which e m a n a t e from
a certain vertex, we see t h a t / c (n, n) _-->n + 1 and h (n, d) ~_ n for 0 ~ d -----n - 1.
W e t e r m i n a t e this section b y giving several applications of 2.6. First we
r e m a r k t h a t when X is a union of closed rays e m a n a t i n g f r o m 0, t h e n p o s X
-- c o n X . I n view of this, the reader will note places in the following discussion
where the word "positive" can be replaced b y '"convex".
 The Generation of Convex Hulls 11

2.11. CorollalT. I[ (X~) is a / a m i l y of closed rays [closed convex cones] each

with vertex 0 and i / p o s U X~ = R ~, then/or each integer d with 0 g d <= n there
exist ]a o/these rays [cones] with j a g m a x (n + 1, 2d) such that the positive hull
o/their union is [contains] a linear subspace o/dimension ~ d.
Pro@ Replace each r a y [boundary ray] of X~ b y a point ~= 0 on it and
apply 2.6.
Before stating a result which is"polar" to 2.11 we introduce the terminology
and state the properties which will be needed. I n this paragraph X [resp. X~]
will always denote a closed convex cone of R r~ with vertex 0. The polar of X
is defined b y
X° . = {y : (y, x) < 0 for all x E X ) .
Then X ° is also a closed convex cone with vertex 0, and X °° = X. X ° is k-
dimensional iff the largest linear subspace contained in X has dimension
n - k. Thus X is a closed half-ray emanating from 0 iff X ° is a closed half-
space with 0 in its boundary. Also, ( Q X~)°= cl pos ( U X°), from which it
follows that 9 : {0} pos(U XO): For proofs of these statements
and a further discussion of this topic the reader is referred to SAXDG~N [36].
The polar version of 2.11 is
2.12. Corollary. I / Z ~ is a [amily of closed hal/spaces in R ~, each having 0
in its boundary [a ]amily o/ closed convex cones in R" with vertex 0] and
i/ Q Z: = {0} then/or each integer e with 0 < e < n there exist k, o/these hall-
spaces [cones] with ke <= m a x ( n + 1, 2 ( n - e)) whose intersection is a linear
subspace [closed convex cone] o/dimension <=e.
Pro@ We will prove the statement in brackets. Since pos( U Z~°)= R n,
b y 2.11 there exist ~ = / ' . - e of these cones with j ~ m a x ( n + 1, 2(n - e)) such

(pos 1 Z
0))0= Z / i s a closed convex cone of dimension < n - [ < e. []
2.12 can be reformulated in terms of hemispheres and great circles on the
surface of the sphere. Let

S~ "= x C R~ +l : z~ x~ - - - 1 ,
1
the surface of the unit sphere in R ~+1. A closed hemisphere of S ~ is a set of the
form {x E S n : (x, w) g 0} where w is a point of R n+~. G is a great circle o / S '~ o/
dimension k iff G is the intersection of S ~ and a k-dimensional linear subspace
of R ~ +1. Now we can state
2.13. Corollary. I / Z ~ is a /amily o/ closed hemispheres o/ S ~ such that
rl z~ = 9 then/or each integer e with 0 ~ e ~ n + 1 there exist lc, o/these closed
O:

hemispheres with ke ~ m a x (n + 2, 2 (n + 1 - e)) whase intersection is contained

in a great circle o/ dimension g e.
Proo/. This is an immediate application in R ~+1 of the unbracketed
portion of 2.12. [3
12 WILLIAM BONNICE and VICTOR KLEE:

§ 3. Types of positive independence

The notion of positive independence has been studied by DAvIs [15] and
McKINN~Y [31], a set X being called positively independent provided for each
x E X, x ~ pos(X ~ {x}). McKINN]~¥ studies also the strong positive independ-
ence of X, this being the condition that 0 $ X and (pos U ) ~ (pos V ) ( {0}
whenever U and V arc disjoint subsets of X. We shall establish some implica-
tions which involve these and related conditions, our principal results (3.4 and
the example after 3.5) solving a problem proposed by McKI~NEY on p. 146
of [31].
For a set X, ~ X will denote the set of all functions ~ on X to [0, ~ [ for
which the set {x E X : ~x =~ 0} of elements involved in ~ is finite ; the cardinality
of this set is called the length of ~ and denoted by 2(~). When X lies in a linear
space and p E pos X ~ {0}, a representation o / p (in terms o / X ) is a function
E ~ X for which p = ~ ~xx; a minimal representation is a representation of
xEX
minimal length, and the length of p is the length of its minimal representations.
Note that although the following conditions of independence are not all
equivalent when phrased in terms of positive combinations, the analogous
conditions for linear combinations are all equivalent to the ordinary condition
of linear independence. (The conditions apply to a set X C L ~ {0}, where L
is a linear space.)
(I) Representations are unique; i.e., if ~ ~xx = ~ fl~x :~ 0 with ~, fl ~ ~ X ,
xEX xEX
then a = ft.
(I') Equivalent representations involve the same elements; i.e., if ~ ~ x
xEX
= X fl~x =4=0 with ~, fl E ~ X , then {x: ~ =~ 0} = {x: fl~,~= 0}.
xEX
(I") Equivalent representations have the same length.
(I"') Representations o/given length are unique; i.e., if ~ ~xx = ~ fl~x=4=O
xEX xEX
with ~, fl E ~ X and 2(a) -- ~((fl), then a = ft.
(I'"') Equivalent representations o] the same length involve the same elements.
(II) I/oc and fl are equivalent representations and ~ is minimal, then ~ ~ ft.
(II') I/o~ and fl are equivalent representations and ~ is minimal, then every
element involved in ~ i8 also involved in ft.
(II") X is strongly Tositively independent.
(III) X is positively independent.
(III') 2Vo proper subset o / X has the same positive hull as X itsel/.
We consider also:
(M) Minimal representations are unique.
(M') Equivalent minimal representations involve the same elements.
3.1. Proposition. Conditions (I), (I'), (I"), (I'"), and (I'"') (expressed in
terms o / ~ s i t i v e combinations) are all equivalent to linear independence.
Proo/. Certain implications being obvious, we show only that (I") and (I'"')
imply linear independence of X. Suppose X is linearly dependent, whence there
 The Generation of Convex Hulls 13

are disjoint n o n e m p t y subsets U and V of X and strictly positive n u m b e r s

~u, ~ such t h a t ~ ~uu = ~ ~vv # 0.
uE U vE V
If (I") holds, t h e n card U = card V and with v0 ~ V and suitable fl > 0
We have
~v0+2:~u=(~v°+~)Vo+ ~ ~ov#0,
uE U vC V ~ {v0)

giving rise to two equivalent representations of different length and contra-

dicting (I"). Thus (I") implies linear independence.
I f ( I ' " ' ) holds, t h e n U and V m u s t be of different cardinality - - say card U <
< card V. L e t W C V with card W = card V - card U. T h e n for suitable
fl > 0 we have

wE W uE U wE W vE V ~ W

giving rise to two equivalent representations of the same length which do not
involve the same elements. This contradicts ( I ' " ' ) and shows t h a t (I'"') implies
linear independence. I~
3.2. L e m m a . The elements involved in a minimal representation are linearly
independent.
Proo]. Suppose X C L, p C L ~ {0}, and a is a minimal representation of p
in t e r m s of X. We wish to establish linear independence of the set
U = (x ~ X : ~ x ~ 0}. Suppose U is linearly dependent, whence there are
n u m b e r s fl~ such t h a t ~ fl~u = 0 and a t least one flu is positive. L e t
uEU
r o = m i n { ~ / f l ~ : u ~ U, fi~ > 0} and let u o (- U be such t h a t ~,/flu. = to. Let
= ~ - r o ft. Then y is also a representation of p in t e r m s of X, and this
contradicts the m i n i m a l i t y of :¢, for ~uo = O. U
3.3. Corollary. Conditions (M) and (M') are equivalent.
3.4. Theorem. For finite X, (M) implies (II').
Proo/. F o r 1 g j ~ k, let S~. k be the s t a t e m e n t t h a t if X is finite and
satisfies (M), a and fl are equivalent representations of length ?" and k respec-
tively, and a is minimal, then {x: ~x ~: 0} C {x: fl~ =~ 0}. Then for each k,
Sa, k is i m m e d i a t e from uniqueness of minimal representations, and to show
t h a t S~.,k is true f o r all j and k we proceed b y " d o w n w a r d " induction on the
first index and " u p w a r d " induction on the second. Specifically, it suffices to
deduce Sj, k (where 1 ~ j < k) from the assumption{ rl kSi,~)~ Sj+I, k.
-- \l~y<k'<
N o w consider :¢ a n d fl as described, with 2(a) = ] a n d 2(fl) = k. There are
subsets U a n d V of X a n d positive n u m b e r s g~ a n d fly such t h a t card U = ],
card V = k, and

uEU uEV

W e wish to prove t h a t U C V. L e t y be a minimal representation of p in t e r m s

of V. If ~(~) < ~(~) it follows from the inductive hypothesis (in fact, from
i ~ k ' fl
< k $i,~') t h a t Y C ( v E V: ~ ~= 0). We may assume, therefore, that
14 W I L L ~ BONNICEand VICTORKLEE:

~(~) = ~(fl), whence fl is also a minimal representation for p in terms of V

and V is linearly independent by 3.2.
For each v E V and ~ E ]0, 1[, let

q~,~ = (~P ÷ ( 1 - ~)v= ( u ~ u ( ~ u ) ÷ ( 1 - ~)v .

Each point q~, ~ is of length g ?"÷ 1 in terms of X. If some q~, ~ has length equal
to j ÷ 1, then the above expression affords a minimal representation of q~, ~ in
terms of X and it follows from S~.+I, k t h a t U ~y {v} C V, whence of course
U ( V. We m a y assume, therefore, t h a t each q~,. is of length ~ j in terms of X,
and will complete the proof b y showing t h a t this assumption leads to a contra-
diction.
Let us fix a sequence (~i) --> 1, and consider an arbitrary v E V. For each
i there is a subset Xiv of X such t h a t cardX~ ~ ~ and q~,v~ posXiv • Since X is
finite there is a set Yv C X such t h a t X/v -- Y~ for infinitely m a n y values of i.
Since the set pos Yv is convex and intersects every line (in particular, the fine
determined b y p and v) in a relatively closed subset of the line, it follows t h a t
P E pos Yr. But card Y~ ~_ ~ and q is the unique minimal representation of p
in terms of X ; hence Y~ = U. Now recalling the definitions of q~, ~ and Y~,
we see t h a t V must be contained in the linear hull of U, an impossibility for V
is linearly independent and card V > card U. [~
3.5. Theorem. For all X, strong positive independence implies uniqueness
o/minimal representations.
Proo/. Suppose X is strongly positively independent and consider minimal
representations a and fl of a point p ~ 0. We wish to show (recalling 3.3) t h a t
(x: ~ > O} = (x: fix > 0}. Let U = {x: ~ > fix} and V = { / : fl~ > ~x). Then

X - fl )u = X (fl - v)v,
uEU vEV

and since U A V = 0 it follows from strong independence of X t h a t these sums

are equal to zero. Hence

p= Z Z
uE U ~E X ~ U vE V xE X ~ V

and since the original representations were minimal it follows t h a t fl~ > 0 for
all u E U and a, > 0 for all v ~ V. This implies the desired conclusion. D
Under the additional assumption t h a t p o s X is a linear subspace, 3.5 was
proved b y I~IcKINNEr (p. 146 of [31]). He then raised the following question:
I f X is a positive basis for a linear space L (i.e., p o s X -- L and X is positively
independent) and minimal representations in terms of X are unique, must X
be strongly positively independent ? For finite X an affirmative answer is
provided b y 3.4 (in view of 3.6). We now show b y an example t h a t the implica-
tion fails when the dimension of L is at least the cardinality of the scalar field;
specifically, we discuss the case of the real field with d i m L ----2 *0. (Our example
depends on the continuum hypothesis, and we have been unable to decide what
happens in t h e real case w h e n d i m L = x0).
 The Generation of Convex Hulls 15

Let E be a real linear space of dimension 2 ~0. T h e n c a r d E = 2 ~° a n d (using

the c o n t i n u u m hypothesis) the set E ~ {0} can be indexed as {xo: 0 G 0 < ffJ},
where Y2 is the first uncountable ordinal and where we require in addition t h a t
x a = x o + x 1 ÷ x 2 while the points x0, xl, and x 2 are linearly independent.
N o w let X~ = {x0, Xl, x2} and having defined X o for a n y 0 let Po denote the set
of all points of the form ~u + fiv for a ~ 0 -<- fl and u, v E Xo (i.e., points of
length =< 2 in terms of Xo). F o r 2 G ~ < ~ , the sets X~ are defined inductively
as follows : let yn be the first point of the set E ~ Uo < ,7Po (i.e., t h a t point x~
of the set which has the smallest index ~); then let z, be the first point of E
which is not in the linear hull of the set (oU<,lXo)w {y~} and let z; = Yn - zn;
finally, set X , = (0<U Xo) W {z~, z',7}. Note t h a t

04= Y~ = x3 = x0 + xl + x2 = z2 + z~ ,
so the set X = U X~ is not strongly positively independent. F r o m the m e t h o d
~<CJ

of construction it is clear t h a t U P,7' ~ E, so to complete the discussion it

~<19

suffices to show t h a t X is positively independent while each point of E admits

at m o s t one representation in terms of two or fewer points of X.
Suppose X is not positively independent, whence there is a first index ~ < £/
for which X n is not positively independent. This implies t h a t one of the points zn
p
and z n is a positive combination of the other together with points of U X o --
say
r
z~ = ~ z , + 5" fl.u,
uEU

where a > 0 < ft, and U C Xe for some ~ < ~1. B u t t h e n

(1 + a)z,j = y, - ~ fl,,u,
uCU

an impossibility in view of the m a n n e r in which z, was chosen. T h u s X is

positively independent.
N o w suppose there is a first ~ such t h a t some point of E ~ {0} admits two
different representations in terms of two or fewer points of Xn. F r o m the choice
of ~ and certain facts of linear independence involving %, z',l and 0~n X,, we
?
see t h a t b o t h of the representations m u s t involve z, or %, t h a t t h e y cannot b o t h
?
involve b o t h z, and z,. and t h a t t h e y cannot b o t h omit the same one of z,
and z;. Thus for suitably chosen ~ < ~, points u and v of X~. positive n u m b e r s
and y, and nonnegative n u m b e r s fl and ~, we have
p
otz, + flu = y z , + (~v ,
or
a z , + flz', = ~,z, + (~v with fl>0
or

~ z . + / % ; = ),z', + ~v with fl>0.

I n the first case we see t h a t
(a + ~)z, = y y , - flu + ~ v ,
16 ~rlLLIAM BON~CE and VmTORKLEE :

whence (recalling tlie manner in which % was chosen) ~ + y = 0 and y, E P~,

contradicting the choice of Yr" I n the second case,
(o~ - fl - ~,)z,~ = - fly,~ ÷ (~v ,

whence y, = (~/fl)v C P~. The third case is like the second, and this completes
the discussion.
When the cardinality of X is unrestricted, uniqueness of minimal represen-
tations does not even imply positive independence of X. (Consider the case
in which X includes exactly one point in each r a y emanating from the origin !)
However, we are able to prove
3.6. Theorem. F o r all X, the conditions (II), (II'), and ( I I " ) are mutually
equivalent, and are also equivalent to the condition that (M) holds/or every subset
o/X.
Proo/. Since strong positive independence is a property of finite character,
it is evident from 3.4 and 3.5 t h a t (II") is equivalent to the condition t h a t (M)
holds for every subset of X. F r o m 3.4 and 3.5 we see also t h a t (II") implies
(II'). Since clearly (II) implies (II') and (II') implies (II"), to complete the
proof it remains only to show t h a t (II') implies (II). Suppose (II') is satisfied
and let T~ denote the statement t h a t (II) holds whenever 2(fl) g k. Then T 1 is
evident and we proceed b y induction. Suppose T k holds and consider equivMent
representations ~ and fl in terms of X, where ~ is minimal and ~(fl) = k ~- 1.
Suppose there exists u such t h a t ~u > flu ; let ~u = /~u and ~ = 0 for x ~ X ~ {u~.
Since 2(fl - ~) = k ~ 1, a - y and fl - ~ are equivalent representations, and
since (a - ~)~ > 0 = (fl - y)~ it follows from T k t h a t ~ - 7 is not minimal.
Thus the p o i n t ( ~ a x x ] - fluu admits a minimal representation (5 ~: a~ - 7.
\ x cx /
But then a and ~ -~ 7 are two minimal representations which are equivalent
but different, an impossibility since we know already t h a t (II') implies (M). [~
I n closin~ we note the obvious.
3.7. Proposition. (I) implies (II). (II) implies (III). (Ill) and (III') are
equivalent.
§ 4. Baryeenters of measures:
Setting and general results
The notion of the barycenter of a measure, generalizing the notion of convex
combination, has been much studied and applied in recent years. The most
important results concern relationships among the formation of convex hulls,
the formation of closed convex hulls, and other hull-formations defined in
terms of barycenters. We describe here a general framework for the study of
such hull-formatious, and obtain a few new results. Some of the notation and
terminology is not for present needs, but will be useful in subsequent work.
All material o/ the present section is essentially known (cf. K I ~ I N - ~ M U L I ~
[30], CHOQVET [12, 13], BAUER [4, 5], and BISHOP-D~ L E ~ v w [6]), b u t is
given here in a form which is especially convenient for our purposes.
Though a study of barycenters based on a "strong" form of vector integral
would surely be worthwhile, our attention here is confined to the common
 The Generation of Convex Hulls 17

"weak" form of the notion of baryccnter. This has the advantage of requiring
integration only for real-valued functions. However, for the treatment of
measures which are only finitely additive and for the integration of unbounded
functions, it is well to describe explicitly an approach to integration which is
suitable for our purposes. (For an alternate approach, see DU~rORD-ScHwARTZ
07].)
A (finitely additive) measure on a set S is a non-negative real-valued
function # such that. dmn # is an algebra of subsets of S (with S E dmn #) and
#(Xu Y)=#X+/tY whenever X, Y E d m n / t with X ~ Y = 0 . The
members of dmn/t are called tt-meazurable, and a function / on S to [ - cQ, c~]
is said to be it-measurable i f f / - 1 ] ~ , fl[ E dmn/t whenever - c ~ < ~ < fl < c~.
For each bounded real-valued function ] on S and for each finite partition
of S into (pairwisc disjoint)/t-measurable sets, we define
j (/, ~ , #) . = ~ (inf/P)/tP ~ ~_~ ( s u p / P ) # P = . f ( ] , ~ , / t ) .
PE~ PE~
The partitions are directed downward by refinement ( ~ ~- ~ iff each member
of .~ lies in a member of .~), and the usual observation of monoteneity shows
the existence of lower and upper integrals
f]d# = lim~,> f ( / , ~ , / t ) g lim~,,> fl(],g~, I.t) = F / d / t .
The bounded function ] is said to be/t-integrable i f f f ] d/x = j - ] d/t, and this
common value is then denoted by f / d/z. When ] is an unbounded function
o n S t o [ - ~ , ~ ] , sE S, and - c ~ < a < fl < c¢, let

[~s= when ~N/s< fl .

( I s <___
Then / will be called /t-integrable iff/~ is/t-integrable for all ~ and fi as described,
and there exists
lim . . . . . ~ - ~ f ]~d/tE [ - ~ , ~ ] ;
the limit is denoted b y f / d/t. Finally, a function ] on S to [ - ~ , ~ ] will be
called/t-summable iff ] is/t-integrable and f ] d/t E R. B y use of refinements
generated b y simple partitions of the interval [inf/S, sup]S], it can be verified
that every bounded /t-measurable ] is/t-summable, whence of course every
unilaterally bounded (i.e. bounded above or below)/t-measurable ] is/t-inte-
grable. When the algebraic operations and the order are defined pointwisc, the
set L(S, #) of all #-summablc real-valued functions on S forms a vector
lattice and the integral ]-~ f ! d/t is a positive linear functional on L(S,/t)
(i.e., / ~ 0 @ f / d/t ~ 0). The reverse problem, that of representing positive
linear functionals in terms of integration, has been studied in a very general
setting b y HEWITT [23] and others.
Now suppose there is given a set E, and a set F of functions on E to [ - oo, ~ ] .
For each set S ( E , let 2 s denote the algebra of all subsets of S and let SF
(resp. S~> denote the smallest algebra (resp. ~-algebra> of subsets of S which
includes all sets of the form S ~ ] - 1 ]~,/~[ for ] E F a n d - c~ < ~ < ~ < c~.
l ~ t h . Ann. 152 2
18 WILLIAM BONNICE and VICTOR KLEE:

With respect to B and F (regarded as fixed throughout this and the next five
paragraphs) we assign barycenters to certain measures. A measure # on a
subset S of E will be called F-admissible iff SF C dmn/t. A point p E E will be
called an F-resultant of the F-admissible measure # iff f fs d # = / p for all
f E F, w h e r e / s denotes the restriction of / to S. Finally, p will be called an
F.barycenter o f / t iif p is an F-resultant of/~ and S -~ 1. In dealing with un-
bounded sets, it is also useful to consider F - c-resultants and F - c-barycenters,
p being an F - c-resultant of the F-admissible measure # on S iff f / s d # - - / p
for all ] E F for which Is is /t-integrable ("c" stands for "candidate"). The
measure # m a y have m a n y F - c-resultants, but if F separates the points
of E a n d / t has an F-resultant p, then p is the unique F - c-resultant of #.
Of course a measure on S m a y fail to have any F - c-resultant. (In the future,
when F is fixed, the prefix " F - " will be "omitted".)
(In the general setting under discussion, it is also of interest to define
resultant~ and barycenters by means of an inequality f ]sd# <=]~ or f ]s d # >=]~
rather t h a n equality, where the inequality (required for all / E F) is chosen
according to the nature of F (cf. BAYER [4]). When F is a linear set of functions,
this is equivalent to the above definition).
Now suppose ~ is a function whose domain is 2~, and such t h a t for each
S C E, ~ S is a (possibly empty) set of admissible unit measures on S, d m n #
being the same for all # E yJS. Thus there is an algebra of sets, D~S, such t h a t
S• C D~S C 2 s and d m n ~ = D~S for all/~ E ~vS. For each S C E, H~S (resp.
HvcS} will denote the set of all barycenters (resp. c-barycentcrs} of measures
# E ~vS. When ~v and ~v' are functions of the sort described, the function
~v A ~v' is defined b y the condition t h a t (for S C E ) # E (~v A ~v')S iff d m n #
--- D~S ~ D~,S and there exist v E ~pS and v' E ~v'S such that/~ is the common
restriction of v and ~' (when DvS = Dv, S , this says (~v A ~v')S = ~ S ~ ~v'S).
We write ~ ~ ~' iff y~ A ~' = v/.
Among the functions yJ of special interest are ¢ and y, where for S C E,
¢ S (resp. ~S} is the set of all finitely additive (resp. countably additive} unit
m e a s u r e s / t on S for which d m n # = SF (resp. d m n # = S~}. Also of interest
are the functions fl, 2, and w, where for S C E, flS (resp. 2S resp. ~oS} is the
set of all # E ¢ S such t h a t /t is carried b y some bounded (resp. finite resp.
countable} subset of S. Thus, for e x a m p l e , / t E flS iff # C F S and there exists
T E dmn~u such t h a t tt T = # S and [ T is, for each / C F, a bounded subset of R.
Note t h a t if yJS includes the unit measures concentrated at the various points
of S, then H ~ S D S and the operation Hv m a y be regarded as a sort of hull-
formation.
For each hull-formation H~, there is a corresponding notion of ~-extreme
:point. Specifically,
pEexVS iff p E S ~ H~(S~{p}).

Since H~S -~ conS, ex~S ~- exS, the set of all extreme points of S in the usual
sense.
 The Generation of Convex Hulls 19

Note t h a t for / E F, f / s d~t depends only on the restriction of/x to Be, and
thus everything could be phrased in terms of functions ~pwith always DvS = 8F.
However, this would be inconvenient, for in m a n y applications one is concerned
specifically with certain countably additive measures whose domains properly
contain S~. I n any case, various problems on the extension of measures arise
(cf. 5.1 and § 6).
$ $ $ $ $ $ $

In the preceding paragraphs, we have described a very general setting for

the study of barycenters and the associated hulls and extreme points. Though
useful in subsequent investigations, this generality is not actually required
for the present study, as we treat only the case in which E is a real linear space
and F is a linear space of linear functionals on E. This makes available the
convenient geometric language of the theory of convexity, and, as can be
shown b y a simple embedding, the restriction is much less drastic t h a n might
at first be imagined. Indeed, suppose U is a set and V is a linear space of real-
valued functions on U, separating the points of U. L e t V* be the space of all
linear functionals on V. For each u E U, let ~u E V* be defined by: ~,, (v) = v (u)
for all v E V, and let E be the linear extension of ~ U in V*. For each v E l~,
let ~ be the linear functional on E defined by: ~ e = e(v) for all e E E, and let
F = ~ V. Then ~ is a biunique transformation of U into the linear space E
and ~ is a biunique linear transformation of V onto the separating linear sub-
space F of E*. For each subset S of U, it is true t h a t (~S)F = {~A : A E 8v}.
If p E U and # is a measure on 8, then # is V-admissible iff the measure # ~-1
is F-admissible and p is the V-resultant of/~ iff ~p is the F-resultant of/x~ -1.
Henceforth, except when the contrary is made explicit, we operate under the
Standing Hypotheses: E is a locally convex Hausdorff linear space and
F = E*, the space of all continuous linear functionals on E. Barycenters and
the associated hull-formations are defined with respect to F.
4.1. Proposition. For all y~ (as described earlier) and/or all S ( E --
H~S C HVcSC el c o n S ,
~ ~o ~ con S C H~ S .
Proo/. The second inclusion is obvious. For the first, consider a point
x ~ E ~ cl c o n s and observe t h a t (by the separation theorem) there exists
/ E F such t h a t / x < iuf/S. Being unilaterally bounded on S, f m u s t be/x-inte-
grable for each/x E v~8. But clearly [x < f [sd# and it follows t h a t x ~ H~8. [~
4.2. Remark. I t the set ~ S o/unit measures is convex, then Hv8 and H~S
are convex subsets o/E.
The following result is close to one given b y BOURBZrXI [10; p. 81]:
4.3. Theorem. If 8E E and cl c o n s is weakly compact, then each admissible
unit measure lx on 8 admits a barycenter in E.
Proo/. For each / E F ~ {0}, let B, .= {x E E : / x = f / s d ) , a closed hyper-
plane in E. We wish to show that/~pfl~ (0} B y # ~), and with C - = cl c o n s weakly
compact, it suffices to show t h a t if ]i . . . . . ]~ E F ~ {0}, then C is intersected
2*
20 WILLIAM B O O . m E a n d VICTOR KLEE:

b y t h e set f"! B h. For each x E E, let T x = {]ix . . . . . ]~x) ~ R n, whence T i s a

i=l
continuous linear transformation of E into R ~. Note that if g E R ~* and a < fl,
then
T S ~ g-~]~, fl[ = T ( S f~ (gT)-~]=, riD,
where g T E F. Thus we can define an admissible unit measure v on T S b y
setting v Y "= # T - 1 Y for all Y E (TS)R~, (an algebra of subsets of TS).
Since//~* has a finite basis it is easily verified that v admits a barycenter z in
R n, with z Ecl c e n T S b y 4.1. But then z E TC, whence there exists x E C with
n
T x = z and it is evident that x C C ~ i~1Bh" ~
The following result of NAMIOKA [32] and BAUER [3] will be used several
times:
4.4. Theorem. Suppose E is a partially ordered linear space with positive cone
C, L is a linear subspace o / E , and [or each x E C there exists y E L with x ~ y.
Then every positive linear/unctional on L (i.e., ~ 0 on L ~ C) can be extended
to a positive linear ]unetional on E (i.e., ~_ 0 on C).
The next result is a slight generalization of a theorem of KREIN-
~MULIA~ [30]. I t shows that for a subset S of E, the relationship between H ~ S
and cl c o n s is closely connected with the representabflity of certain positive
linear functionals by means of measures from vS.
4.5. Theorem. Suppose E is a real linear space, F is a linear subspace o] E*,
S C E, F' is the set o~ all/unctionals o/the ]orm [ + ¢~/or / E F and ~ E R, and
F ' S • = {['s:]'E F'}. Suppose G is a linear space o[ reaLvalued ]unctions on S
such that F ' s C G and [or each g E G there exists [' E F' with g ~ /'~. Let P denote
the set o] all positive linear [unctionals p on G such that pe = 1, where e is the
]unction identically 1 on S. Then the closed convex hull o / S [or the topology ~(E,F)
is identical with the set K o/all x E E such that there exists p E P h a v i n g / x --- P/s
/or all [ E F.
Proo/. Let C ' = cl conS and suppose there exists a point x E K ~ C. B y
the separation theorem, there exists ] E F such t h a t inf]C > / x . And with
x E K, there exists p E P such t h a t / x = P / s for all ] E F. Then
[ x = P/s ~-- p(ird/S)e = i n f / S >= inf/C > [ x ,
and the contradiction shows t h a t K C C.
To complete the proof it suffices to show that for each x E E ~ K there
exists'] E F having [x < inf~K. We claim first that there exists ]' E F' such that
]'x < 0 but /~ _-> 0. Suppose the contrary. Then whenever ]', [" E F ' with
]~ = / ~ we have (/' - / " ) s :> 0 and (]" - [')s ~- 0, whence (/' - / " ) x ~_ 0 and
([" - ]')x ~_ O. We may then define a positive linear functional q on F's by
specifying that q / ' s ' = f x for all ]' E F'. Since each member of O is majorized
by some member of F~, the Namioka-Bauer theorem 4.4 implies that q can be
extended to a positive linear functional p on G, with pe = qe = 1 and conse-
quently p E P. But P/s = q f s ' = [x for all /E F, so x E K, a contradiction
implying the existence of ]'E/~' such that ]'x < 0 but ]~ ~ 0. Of course
 The Generation of Convex Hulls 21

/' = ] + a for some ] E F and a E R, and thus ]x < inf/S. Now consider an
arbitrary y E K. B y the definition of K there exists p E P such t h a t ]y = P/s
for all ] E F, and hence
/ x < i n f / S = p ( i n f / S ) e < P/s .... / Y .
Thus ]x < i n f ] K and the proof of 4.5 is complete. [~

§ 5. Finitely additive measures

The following remark shows t h a t there are no extension problems when we
are concerned only with finitely additive measures.
5.1. Proposition. I] # is a finitely additive unit measure on a set S, ~ can be
extended to a finitely additive measure whose domain is 2 S.
Proo]. L e t E be the linear space m S of all-bounded real-valued functions
on S, C the positive cone in E, and L the subspaee consisting of all/x-integrabie
functions. For each x E L, let ]x = f x dlx. Then ] is a positive hnear functional
on L, and since L includes all the constant functions it follows from 4.4 t h a t ]
can be extended to a positive linear functional T on E. By a theorem of HmDE-
BRANDT [25] and FICHTENHOLTZ-KA~TOROV~TCI~[18], there is a finitely
additive unit measure fi with domain 2 ~ such that ]y = f y d/2 for all y E E.
Clearly fi is the desired extension of/x.
I t follows from 5.1 t h a t always H ~ S = H ¢. S, where ¢ . S is the set of all
finitely additive unit measures with domain 2 s. Our main result on the hull-
formation H~ is
5.2. Theorem. For all S C E,
con clS C H ¢ e l S - H ~ S C H~(H~'~) = ct c o n S ,
wqth H ~ S -- el con S when S is bounded.
Proo]. B y the basic separation theorem for convex sets, the closed convex
hull of S in the given topology for E is the same as t h a t for the topology a (E,F).
Thus when S is bounded, the fact t h a t H ~ S = cl c o n s follows from 4.5 (with
G = m S ) in conjunction with the representation theorem of I~IILDEBRANDT[25].
I n continuing the proof of 5.2, we shall use freely and without explicit reference
the fact (5.1) t h a t H ~ = H ~'. From 4.1 it follows t h a t for all S,
con cl S C H ¢ clS D H ~ S C H~ (H ¢ S) C el con S .
To complete the proof it suffices to show t h a t H ¢ e l X C H ~ X for all X C E,
for then
cl c o n s C c l H ¢ S C H ¢ ( e l H ¢ S ) C H ~ ( H ¢ S ) •
Now consider a set X C E and a point p E H ~ clX. There exists ~t E ¢ " c l X
such t h a t p is the barycenter of ~u. L e t H be the set. of all functions on E of the
form g + y f o r / E F , {a, fl, y } C R , and a ~ fi, let H x = { h x : h E H } , and
let J be the set of all functions on X of the f o r m / x + u for ] E F and u E m X.
For each h E H, let ~h x = f h¢l:rd~, properly defining the function ~ on H x
for by continuity hcl x is determined b y h x. Then ~ is a positive linear functional
on H x and hence b y 4.4 it can be extended to a positive linear functional ~/on H.
22 WILLIA~Bo~IC~ and VICTORK ~ z :

Since clearly ~ 1 = 1, Hfldebrandt's theorem guarantees the existence of a

unit measure vE ¢ " S such t h a t ~ u = f u d v for all u E r e X . For each [ E F
we have
f [ x dv = lim . . . . . a~+ ~ f [gx dv = l i m ~ / ~ x

where of course the last equality comes from the fact t h a t p is the barycenter
of/~. But then p is also the barycenter of v, whence p E H ¢ X and the proof
of 5.2 is complete. [l
The following result shows t h a t the equality H ~ S = cl c o n s is valid for a
large class of unbounded sets S.
5.3. Theorem. Suppose K is a finite.dimensional closed convex subset o / E ,
and K contains no line. Then the/ollowing/our assertions are equivalent:
(i) K = con c t e x K ;
(if) c l b o n S = K ~ tJ cleonB = K;
bounded B C S
(iii) cl c o n s = K ~ H~ S = K ;
(iv) cl con S -- K ~ H ¢ S = K.
Proo/. Suppose first t h a t (i) holds and K = cl conS. Then clS D cl e x K by
a result of KLEE [27], and hence K = con clS. Each point p of K has the form
eiPt with p~ EelS, ai ~ 0, and ~ ~t -~ 1. For each i, let p}, p~ . . . . . be a
1 1
sequence in S converging to p~, and let B = {p~ : i, ~ C [1, ~[}. Since p E clconB
and B is a bounded subset of S, we conclude t h a t (i) implies (ii).
I t follows from 5.2 t h a t (ii) implies (iii) (with no restriction here on dimen-
sionality), for/-/~' B = el eon B when B is bounded. Obviously (iii) implies (iv).
Now suppose (iv) holds and assume without loss of generality t h a t E is the
affine extension of K, whence K includes an interior point q. Let C be the closed
convex cone which is the union of all rays in K emanating from q, and let
S = C ~J el e x K . B y a theorem of KImE [27], K = c o n ( e x K ~.J rexK), where
r e x K is the union of all extremal rays of K. Since the endpoint of an extremal
r a y is an extreme point, it follows easily t h a t K -- cl conS, whence (by iv)
K = H e S . To establish (i) it suffices to show t h a t if ~ is an extremal ray of K
having endpoint u, and v ~ ~ ~ {u},-then there exists w E ~ (~ con et e x K such
t h a t v E [u, w].
B y maother result of KL~E [27], there exists ] E F ~ {0} such t h a t for all
:~ E R, the set K ~ / - 1 ] _ c¢, ~] xs compact. L e t X = S A ] - 1 ] _ ~ , ]v[, and for
each ~ / v let Y ~ = S ~ / - X [ / v , o : ] and Z ~ = S ~ / - 1 ] ~ , c ¢ [ . Of course
S = X w Y~, ~ Z~.
B y hypothesis, v is the barycenter of a measure ~ E ¢ " S. Let } denote the
restriction of ;u to 2 x, and for each ~ _~ ]v let ~ and ~a denote the restrictions
of ]~ to 2 r~ and 2 z~ respectively. B y 4,3 the measures } and ~ have resultants x
and ya respectively, and then from the relevant definitions it follows t h a t ~
admits a resultant z~ with v = x + y~ + z~. Alternatively, we m a y write
= (~_~)~' + (~ Y~)y~ + (~z~)z'~,
 The Generation of Convex Hulls 23

where x' is an arbitrary point of K if/~X = 0, x' = (*/#X)x when # X ~= 0, and

the points y~ and z~ are obtained in a similar manner. Note t h a t {x', y'~, z~} C K
in any case, since K is closed and convex and (for example) when tu X :4= 0,
x' is the barycenter of the unit measure (*/#X) ~ E ¢ ' X.
The discussion now splits into two cases:
Case I: For some a _ / v , ~ Y~ > O. To treat this case, we consider the
affine projection z of E onto the hyperptane {p E E : / p =: ]u} which sends
into u. Since ~ is an extremal r a y of K, the point ~ v ( = u ) must be an extreme
point of ~ K .
But
~v = (#X)~x" + (~ Y~)~y~ + (#Z~)~z~
with
~ X + /~Y, + ~ Z ~ = I ,
and since/~ Y~ > 0 while a v ( e x z K , it follows t h a t ay~ = 7ev. Thus Q includes
the barycenter y~ of the measure (1/# Y ~ ) ~ and from the definition of Y~ it
follows t h a t v C [u, y~].
Further, we have
Y~ C H ~ Y~ -- con Y~,
where the equality follows from compactness and finite-dimensionality of Y,
and hence
y~ ~ con Y~ ~ b d r y K C con cl e x K ,
where the inclusion is justified by the definition of Y~ in conjunction with the
fact t h a t C C int K. This completes the discussion of Case I.
Case I I . For all ~ ~ /v,/~ Y= = 0. Then ~Z~ = 1 - / ~ X and
v = (~X)x' + (1 - ~ x ) z ' ,
whence
/v >= (/~X) i n f / K + (I - # X ) ~
for all ~ ~ /v. This implies t h a t # X = 1, whence v = x', and the reasoning of
the preceding paragraph shows t h a t x' C con cl ex K. q

§ 6. Countably additive measures

Recall t h a t S~ is the a-algebra generated b y sets of the form S ~ / - 1 ]a, fl[
for / E F and ~, fl C R, and HrS is the set of all barycenters of countably
additive unit measures/z for which dmn/z = S~.
6.1. Theorem. For each S ( E,
c o n s C H~S C et c o n s ,
with H ' S = c o n s when E is finite-dimensional and HvS = cl c o n s when S is
weakly compact.
Proo]. The inclusions are immediate from 4.1. The finite-dimensional
equality was established b y RIC~Tm~ [33] and RUBr~-W~SL~R [35], and is
proved in stronger form below (6.2 and 6.3). T h a t HvS = el c o n s when S is
compact has been established by To~I~X [39], BOURBAK~ [10], H ~ . ~ ' ~ -
SAVA~ [24], and C~OQU~T [12]. I t follows at once from 4.5 (taking for G the
24 WILLIAM BONNICE a n d VICTOR K L E E :

space of all weakly continuous real functions on S) in conjunction with the

Riesz representation theorem (e.g., § 56 of HALMOS [21]).
F o r S C E and x E E, we shall s a y t h a t x is weakly exposed in E ~ S iff
there is a finite sequence (E~)~ of closed flats such t h a t E o ~= E, x E E,~ C E ~ S,
and for each i, S f~ E~_ 1 lies in a closed halfspace of E l - 1 which has E i as its
bounding hyperplane relative t o E l _ i. N o t e t h a t if E is finite-dimensional and
x E E ~ conS, the basic separation t h e o r e m for convex sets implies t h a t S
lies in a closed halfspaee whose bounding hyperplane E 1 includes x. Similarly,
x can be separated from c o n ( S ~ El) b y a hyperplane E 2 in El, and continuing
in this w a y we see t h a t x is weakly exposed in E ~ con S. Thus the result of
RuBI~-WwSLER [35], incorporated in 6.1, is a corollary of
6.2. Theorem. Suppose S C E, x is weakly exposed in E ~ S, and x is the
barycenter o/ an admissible measure t~ on S. Then there exists a closed fiat J o/
finite deficiency in E and an increasing sequence (Qi)1 ~ o/closed hal/spaces in
oo

J ~ {x} such that x E cl y ( S ~ Qi) and ~ U (S f~ Qi) > 0 although I~ ( S ~ Qi) = 0

/or each i. I n particular, # is not countably additive.
Proo/. Let E o. . . . . E n be as described above and let m be such t h a t
# ( S • Era) = 1 b u t / ~ ( S ~ E ~ + I ) < 1. L e t J -- Era, let Q be the closed half-
space in J which contains S ~ E ~ a n d is b o u n d e d b y E ~ + 1, a n d let (Qi)~ be
an increasing sequence of closed halfspaces in J whose union is Q ~ E ~ + i. T h e n

/~ U (Sf~ Qi) = # ( S ~ (Q ~ Era+l)) = 1 - t ~ ( S ~ E,~+I) > O.

There exist/1 . . . . ,/m+lE F ~ {0}, ~ . . . . . ~,,+1~ R, and ~i > 0 such t h a t
m
E m = ~/Tlc~j, Era+l= E ~ ~ / ~ 1 ~ + 1, Q = E ~ ~ / ~ 1 [a,~+l, ¢c[,
and
Q~ = E ~ / ~ 1 [ ~ m + l ÷ ei, ~ [ .
B u t then
a , . + , = Ira+Ix = f / s ~ ~ d~ > a ~ ÷ l ~ ( Z ~ (Q ~ Q,)) + (~,~÷~ + ~ ) ~ Q , ,
whence ~uQ~ = 0.
Our n e x t result was proved in a different w a y b y RICHTER [33] (with
m ~ n only when X is an are). W e are indebted t o R O ~ D P:zK~ for calling
our attention to Riehter's paper, which contains other interesting results
related t o 6.1--6.3.
6.3. Theorem. Suppose i~ is a probability measure on a set X (i.e., a countably
additive unit measure defined over a a,algebra o/subsets o / X ) and ~ , . . . , ~n are
I~.summable real,valued/unctions on X . T h e n / o r some m ~ n + 1, there exist m
points x~ . . . . . x~ o / X and positive numbers ~ . . . . , ~,~ with ~ ~ =: 1 such that
1
m

f ~dl~= ~ ~¢~(x~) /or 1 ~ i ~ n.

t=1

This holds/or some m ~ n i / X is a topological space, all the ~ ' s are continuous,
and
 The Generation of Convex Hulls 25

X has at most n components

or

X is compact and/or each/unction ~ = ~ fit ~i

i=1
on X (with fit ~ R), the image ~ X is connected.
Proo/. For each x ~ X let Tx==(~l x . . . . . ~nx) E R ~. Let S=: T X C R n
and F = (Rn) *. Note t h a t if / E F and ~, fl ~ R then
T - I ( S f-~/-1]~, ~ [ ) ~_ ( / T ) - l j ~ z , ~ [ ,

and since ] T is a linear combination of the ~i's, / T must be #-summable and

( / T ) - I ] ~ , / 5 [ ( dmn/z. Since dmn/~ is a a-algebra, it follows t h a t
d m n # ) { T - 1 A : A ~ S~,}
and a countably additive unit, measure v on S is obtained b y defining
vA=#(T-1A) forall A~S~,.
Let 351 ~(n be the coordinate functionals on R n, and let p E R n with v~p
. . . . .

= f ~ d/z. Then Tip = f v~s dv and it follows t h a t / p = f / s dv for all f E F.

Thus p is the barycenter of v, whence p E c o n s by 6.1 and p E c o r ~ + l S by
Carath6odory's theorem. Under the supplementary hypotheses of 6.3, the set S
has at most n components or is compact, and convexly connected, whence
m m
p ~ connS by 1.5. Thus p has the form ~ a~q~ for q~ E S, ~ > 0 with ~ ~ j = 1,
1~1 ]=1
and m as described in 6.3. Choosing xj ~ X with Tx~ = qj, we see t h a t f ~i d/~
~n
= ~ a j ~ ( x j ) f o r l ~ i ~ n.[]
i=1
From more careful use of 1.5 and 6.2, sharper forms of 6.3 can be derived.
However, these are rather complicated to state and perhaps only of marginal
interest.
We have observed t h a t H¢(H~S) = H ¢ S for all bounded S (5.2) and also
for certain (but not all) unbounded S (5.3). F r o m 6.1 it follows that H~(H~S)
= H r S whenever S is weakly compact or S lies in a finite-dimensional linear
subspace of E. I n the remainder of this section, we shall describe an unbounded
countable set S for which Hv(HvS) ~ HrS. A tool in the description is the
following result of SC~AVDER [37] and GOSSlEXVX-PArV [20], for which we
indicate a short proof.
6.4. Proposition. Suppose E is a metric linear space, X C E, p ~ E, (ai)~
is a sequence o/nonzero real numbers, (~/Z¢l)p ~ clX, and 0 ~ int clX. Then there
is a sequence (x~)~ o/ distinct points o / X such that p = ~ a~x e
1
Proo/. Let @ be an invariant metric for E. Since ~ p ~ clX, there is a
sequence (Yt)~ of points of X such t h a t y~-~ a ~ p , whence g~y~-~p and
~(p- ~ y ~ ) - ~ 0 . Thus there is a m e m b e r x~ of the set {y~}~ such t h a t
@(p, a~x~) < 2 -~ and a g l ( p _ ~x~) ~ cl(X ~ {x~}). Repetition of this reasoning
produced a point x~ of X ~ {x~} such t h a t O(p - ~x~, ~x~) < 2 -~ (whence
26 WIILIAM Bo~mc~ a n d VICTOR K L E E :

2
e(P, axxl + a, x2) < 2-*) and a ~ ( p - ~ alxi) E c l ( Z ~ {xl, x2}). I t is clear
1
t h a t the proof can be completed b y a continuation of this process.
A (Schauder)basis for a Banach space E is a sequence bl, b 2. . . . of points
of E such t h a t for each x E E there is a unique sequence (xi)~ of real numbers

for which lim a_~ oo xibi : X, the convergence being for the norm topology of
i
E. An equivalent requirement is obtained when norm-convergence is replaced by
weak convergence. The basis is said to be unconditional iff for each x E E, the
series ~ x ~bi converges unconditionally to x; otherwise the basis is conditional.
1
Many conditional bases are known; in particular, the space C [0, 1 ] admits a
basis and all its bases are conditional. (Results on bases which are used here
without proof or exact reference m a y all be found in BANACH [2], KARLIN [26],
or DA:~ [16].) 1)
6.5. Proposition. I f (bi)~ is a conditional basis for a Banaeh space E, and
S = {-b~}~ ~J {bi}~, then H v S lies in a proper linear subspace of E.
Proof. Since H r S is convex and symmetric about 0, the set tJ :¢HrS
a~0
is a linear subspace of E. Thus if 6.5 fails, each point x E E is a positive multiple
of a point of H ~ S and hence is the resultant of some countably additive
measure/~x with dmn/~, = S~. Since the coordinate funetionals y~ I Y E E are
linear and continuous, it is evident t h a t each onepointed subset of S is a member
of S~ and hence is #~-measurable. But S is countable and #~ is countably
additive, so there exist functions ~x on {-bi} ~ to [0, ~ [ and ~ on {b~}~ to
[0, c~ [ such t h a t for each T E S~,
#~ T : ~, ~ ( - b~) ÷ Z ~ ( b i ) "
--blE T b~E T
For all x E E and ] E F we have / x = f/s d~, and since the definition of
integral is such t h a t summability implies absolute summability, we conclude
that

i=1 i=1
where both series converge absolutely and hence unconditionally. Then of
course

/ x = ~ [rl~(bi)- ~ ( - b i ) ] f b i ,
i=l
where again the convergence is unconditional. I t is thus evident that for all i,
Xi = ~ x ( b i ) - ~x(- hi),
and the sequence (bi)~ is in fact an unconditional basis for E, contradicting
the initial hypothesis and completing the proof. [~
1) Added in proof: G~a.~rf [40] has proved that if E is a separable infinite-dimensional
Banach space that admits a basis, then some closed linear subspaee of E admits a conditional
basis. We conjecture that E itself must admit a conditional basis, but are unable to prove
this.
 The Generation of Convex Hulls 27

6.6. Proposition. Suppose U is the open unit cell o/ a Banach space E and
(bi) ~ is a conditional ba.sis /or E. Then there exists a sequence (a~)~ in tg such
that whenever fli ~-- o~i and S ~- ( - flibi}~ '~J (flib~}~, then U C Hr(HrS) but
U ~ H~'S.
Proo/. Whenever /~i > 0, the sequence (flibi)~ iS also a conditional basis
for E, whence 6.5 implies that U ([ HvS. I t suffices, therefore, to produce a
sequence (~i)~ of positive numbers such that if si = aibi and S ==( - si}~ ~J
~, (si}~, then (conS) ~ U is dense in U, for then

U C H ~ (H~S) C H~ (H~S),

where the first inclusion follows from 6.4 and where eo (resp. 2) here indicates
eountably additive unit measures carried by countable (resp. finite) sets.
For each n, let Zn be the canonical projection of E onto its subspace
En = 2 Rbi, whence :~,~x= ~
1 l
xtb~. Then E can be renormed (preserving its
topology) by a norm ( ( ) ) such that 2 ( ( ) ) ~ II II and such that each
projection ~ is of norm 1 relative to (( )). Let V be the new closed unit cell
{ x C E : ( ( x ) ) < 1}, whence U ( ~ V . For each n, let U~= U ~ E n and Vn
= V~E~.
Now in choosing (~i)~, we begin with al such that ( ( ~ l b l ) ) = 1. When
~1. . . . . aj have been chosen, C/will denote the interior of the set con ( ( - ~ibi) ~
(~bi)~) relative to the containing subspace E.. Suppose gl . . . . . ~k have
been chosen in such a way that 1 V~C Ci for 1 -< i -<- k. (Such a choice has
already been effected for 1~:= 1.) Since z~+l is of norm 1 relative to (( )),
it follows that gk+l Vk+l= Vk and hence 1 V k + I C C k _~ Rbk+l. It is easily
verified that
Ck ~- Rbk+1 U Di, where D i . = c o n ( C ~ ' ~ ] - i b k + l , ibk+l[).
~i~l

But D ~ D e c ' " , each set Di is open relative to E~+~, and the set ~ V~+~
is compact, whence there exists i~+~ such that 1 V~+~QD~+. We then set
~ + 1 = i~+~, whence ~ V~+~ Q C~+t and the choice of the ~i's is carried on b y
mathematical induction. With S = ( - ~¢~bi}~ ~ (~bi}~, it is now clear t h a t

whence {con S) f~ U is dense in U and the proof is complete. [~

Obviously the gi's in 6.6 can be chosen so that the set S is unbounded.
We suspect that S must be unbounded, and indeed that most "natural" hull-
formations Hv act idempotently on wide classes of bounded sets. In particular,
we conjecture that H , ( H r S ) = H v S whenever S is a bounded Borel set in a
separable Banach space E. Under the supplementary hypothesis that E* is
also (norm-) separable, this has been proved elsewhere (K~.E [29]).
Finally, we refer to BJSRC~ [7] for some results on the hull-formation Hv
which are close to the spirit of our work.
28 WILLIAMBONNICEand V~CTORKLEE :

Bibliography
[1] ABE, Y O S ~ I , T o ~ o KU~OTA and HAJIMIY YO~EOtrC~: Some properties of a
set of points in Euclidean space. Kodai Math. Seminar Reports 117--119 (1950).
[2] BANACH,S.: Th6orie des op6rations lin~aires. Warsaw 1932.
[3] BAUER, HEINZ: ~ber die Fortsetzung positiver Linearformen. Bayer. Akad. Wiss.
Math.-Nat. KI. S.-B. 1957, 177--190 (1958).
[4] - - ~flovseher Rand und Dirichletsehes Problem. Ann. Inst. Fourier {Grenoble) 11,
89--134 (1961).
[ 5 ] - Axiomatisehe Behandlung des ])irichletschen Problems fiir elliptische und
parabolischc Differentialgleichungen. Math. Ann. 146, 1--59 {1962).
[6] BIsHoP, E., and K. DE LEEVW: The representation of linear funetionals by measures
on sets of extreme points. Ann. Inst. Fourier {Grenoble) 9, 305---331 (1959).
[7] BJS~K, G 6 ~ : The set of extreme points of a compact convex set. Ark. Mat. 3,
463--468 (1958).
[8] BequESts, F., and W. FE~CH~L: Theorie der konvexen Kbrper. Berlin: Springer
1934. (Reprint Chelsea, New York 1948.)
[9] Boo,mE, W. E.: A generalization of a theorem of Carath~odory. Am. Math. Soc.
Notices 8, 252--253 (1961).
[10] BOU~AF~I, N.: Integration. Paris: Hermann (1952), Chaps. I - - I V , A. S. I. @ 1175,
[11] CA~THkODOI~Y,C.: ~ber den Variabilit~tsbereich der Koeffizienten yon Potenz-
reihen, die gegebene Werte nicht annehmen. Math. Ann. 64, 95--115 (1907).
[12] CHOqVET, GUSTAV~: Theory of capacities. Ann. Inst. Fourier (Grenoble) ~, 131--295
(1955).
[13] - - Le th6or~me de repr4sent~tion int~grale dans les ensembles convexes compacts.
Ann. Inst. Fourier (Grenoble) 10, 333--344 (1960).
[14] DA~ZER, L., B. GRi~NBAVr~and V. KLEE: Helly's theorem and its relatives. Pro-
ceedings of Symposia in Pure Mathematics Vol. 7, "Convexity", Am. Math. See.
(1962), 101~180.
[15] DAvis, CHANDLER: Theory of positive linear dependence. Am. J. Math. 76, 733--746
(1954).
[16] DAY, M.tHLONM.: Normed linear spaces. Berlin-Gbttingen-I{eidelberg: Springer
1958.
[17] DUNFORD,NELSON, and JACOB T. SCHWARTZ: Linear operators. Part I: General
theory. New York: Interscience Pub. 1958.
[18] l~cnT~.m~OLZ,G., et L. KANTOROVICE: Sur les operations dana l'espace des functions
born6es. Studia Math. 5, 69--98 (1934).
[19] GALE, DAVII): Linear combinations of vectors with non-negative coefficients. Am.
Math. Monthly ~9, 46---47 (1952).
[20] GossIEavx, A~E-M_tmE, et G~OROES Px~Y: Un th~or~me sur les espaces du
type F. Bull. See. Roy. Sci. Liege 1~, 146--150 {1944).
[21] H~LMOS, PAUL R. : Measure theory. New York: Van Nostrand 1950.
[22] H ~ R , O., and H. RA~sTRbM: A generalization of a theorem of Fenchel. Prec.
Am. Math. See. ~, 589--593 (1951).
[23] HEwrrr, EDWIN: Integral representation of certain linear functionals. Ark. Mat. 2,
269--282 (1952).
[24] - - , and L. J. SAvior: Symmetric measures on cartesian produc~s. Trans. Am. Math.
Soc. 80, 470---501 (1955).
[25] HILDEBRANDT,T . H . : On bounded linear functional operations. Trans..Am. Math.
Soe. ~ , 868--875 (1934).
[26] K~_-_~, S.: Bases in Banaeh spaces. Duke Math. J. 1~, 971--985 (1948).
[27] KLan, VICTOR: Extremal structure of convex sets. Arch. Math. 8, 234--240 (1957).
[28] ~ The generation of affine hulls. Acta Math. Szeged 24 (1963), to appear.
[29] ~ Idempotency of the hull-formation Hr. Z. Wahrscheinlichkeitstheorie l, 258--262
(1963).
 The Generation of Convex Hulls 29

[30] KR~.IN,M., and V. ~UI~AN: On regularly convex sets in the space conjugate to a
Banaeh space. Ann. Math. (2) 41, 556--583 (1940).
[31] McKINNEr, R.L.: Positive bases for linear spaces. Trans. Am. Math. Soc. 105,
131--148 (1962).
[32] NA~IOXA,ISAAC: Partially ordered linear topological spaces. Mere. Am. Math. Soc.
24, 50 pp. (1957).
[33] RICHTER,HAWS: Parameterfreie Absehiitzung und Reatisierung yon Erwartungs-
werten. B1. Deut. Ges. Vers.-Math. 8, 147--162 (1957).
[34] ROBINSON,C.V.: Spherical theorems of Helly type and congruence indices of
spherical caps. _Am.J. Math. 64, 260--272 (1942).
[35] RuBriC, R~RMAN, and OSCARW~sL~: A note on convexity in Euclidean n-space.
Proc. Am. Math. Soe. 9, 522--523 (1958).
[36] SANDGRgN,LSNNART: On convex cones. Math. Scan& 2, 19--28 (1954).
[37] SCHAVDER,J.: ~ber die Umkehrung linearer stetiger Funktionaloperationen.
Studia Math. 2, 1--6 (1930).
[38] STEINITZ,E.: Bedingt konvergente Reihen und konvexe Systeme I - - I I - - I I I .
J. reine angew. Math. 148, 128--175 (1913); 144, 1---40 (1914); 146, 1--52 (1916).
[39] TOMITA,MINORU: On the regularly convex hull of a set in a conjugate Banach
space. Math. J. Okayama Univ. 3, 143--145 (1954).
[40] GURAmI, V.I.: Inclinations of subspaces and conditional bases in Banaeh spaces.
Soviet Math. 3, 1028--1031 (1962).

(Received October1, 1962)