Вы находитесь на странице: 1из 9

Reprinted from Proceedings of the Symposium on Pure Mathematics, (1963) "The Support

Functionals of a Convex Set," E Bishop and R R Phelps, Volume 7, pp. 27-35, by permission
of the American Mathematical Society.

THE SUPPORT FUNCTIONALS OF A CONVEX SET


BY

ERRETT BISHOP AND R. R. PHELPS


The following well-known separation theorem is basic to the considerations
of this paper.
SEPARATION THEOREM. Suppose that A and B are convex subsets of a real
Rausdorff topological vector space E, and that the interior of B is nonempty
Selected Papers of Errett Bishop Downloaded from www.worldscientific.com

and disjoint from A. Then A and B can be separated by a hyperplane, that


by LA TROBE UNIVERSITY on 04/06/17. For personal use only.

is, there exists a continuous linear functional f $ 0 on E such that sup f{A) ^
inf/(B).
This theorem is a geometric version of the Hahn-Banach theorem. Its proof
can be found in any of several texts, for instance in [3, p. 417]. An immediate
corollary is the following support theorem,
If C is a convex subset of a real Hausdorff topological vector space E, if x is
a point in the boundary of C, and if the interior of C is nonempty, then there
exists a hyperplane which supports C at xf that is, there exists a continuous
linear functional / ^ 0 on E such that f(x) = sup/(C). We refer to such a
functional / as a support functional of C, and x is called a support point. Note
that if / is a support functional of C then every positive multiple of / is also
a support functional of C.
The assumption that C has interior points is a strong one, but some con­
dition is indispensable to the validity of the support theorem. Indeed,
V. L. Klee has shown [6] that there exists a bounded closed convex subset
of a dense subspace of a Hilbert space which has no support points. In the
same paper Klee asked whether every bounded closed convex set C in a Banach
space has at least one support point.
In this paper we answer Klee's question affirmatively: We show that the
support points of C are actually dense in the boundary of C. This is shown
to be true even if C is not bounded. Still assuming that E is a Banach space,
we then prove (if C is bounded) that the support functionals of C are dense
in the dual space E*. If C is not necessarily bounded, we show more gener­
ally that for each / in E* which is bounded on C and each e > 0 there exists
a support functional g of C with | | / — g\\ < e. In fact it is shown that g can
be chosen to strictly separate C from any bounded set X which is strictly
separated from C by / . We also show that every hyperplane which intersects
the boundary of C contains a support point of C. Examples are given to
show that these theorems fail in certain more general situations.
The methods of this paper derive from a previous paper [1], in which a
proof was indicated of the fact that the set of support functionals of a bounded
closed convex set in a Banach space is dense in the dual space. By extend­
ing and simplifying the method of [1], we have been able to improve this
27

293
28 ERRETT BISHOP AND R. R. PHELPS

result and to obtain proofs of the related theorems mentioned above.


Although all our theorems are stated and proved for spaces over the real
field, they may easily be formulated and extended to spaces over the complex
field by applying them to the underlying real space (obtained by restricting
multiplication to real scalars) and to the real parts of complex linear function­
a l . These formulations are analogous to that given in [3] for the separation
theorem.
Throughout this paper we will restrict our attention to normed spaces E
and the convex sets under consideration will be assumed to be proper and
nonempty. We write U for {x: \\x\\ ^ 1}; if / e £ * , then | | / | | is defined to
be supf(U).
The proofs of the existence of support points and support functionals are
based upon showing the existence of certain support cones for a given closed
Selected Papers of Errett Bishop Downloaded from www.worldscientific.com

convex set. We will say that a subset K of £ is a convex cone if /if is a


convex set and Ay e K whenever y e K and A ^ 0. If X is a set containing the
by LA TROBE UNIVERSITY on 04/06/17. For personal use only.

point Xo, and if K is a convex cone such that K 4- x0 is disjoint from X ~ M,


then we say that K + x0 supports X at x0. Suppose, now, that K has non­
empty interior, that C is convex, and that K 4- x0 supports C at x0. Then,
by the separation theorem, there exists a nontrivial g in E* such that
sup#(C) ^ infg(K-\- Xo). It is easily verified (since x0 is in both K + x0 and
C) that supg{C)=g(x0) = Mg(K + JC0), i.e., g supports C at x0. Thus, to
show the existence of support points and support functionals for C it suffices
to find support cones of C which have interior points. The cones which we
will use for this purpose are all of the following type:

DEFINITION. It / is an element of E* of norm one and if k > 0, then K(f, k) =


{*: | | * | | :£*/(*)}.
Clearly, K(f% k) is a closed convex cone; furthermore, if k > 1, then the
interior of K(/, k) is nonempty. To see this, choose x in E such that 11 x 11 = 1
and f(x) > k~l. Since / and the norm are continuous, and since || JC|| < kf(x)9
it follows that this (strict) inequality is true for all points in a neighborhood
of x.
The fact that closed convex subsets of a Banach space admit support cones
is a consequence of the following lemma.
LEMMA 1. Suppose that X is a complete subset of a normed linear space E,
that f in E* is of norm one and is bounded on X, and that k > 0. If zeX,
then there exists a point x0 in X such that x0 e K(f, k) + z and K(f, k) + x0
supports X at xQ.
PROOF. We partially order the set X by means of K; that is, x > y means
x — y 6 K. This, of course, is equivalent to saying that || x — y || ^ kf(x — y).
It is easily seen that if there exists a maximal element x0 in X, then K + x0
supports X at x0. To obtain the conclusion that such an x0 exists for which
Xo > z, it suffices to apply Zorn's lemma to the set Z of those x in X for
which x > z. Suppose, then, that W is a totally ordered subset of Z. The
set {f(x):xe W} is a bounded monotonic net of real numbers (using W as

294
THE SUPPORT FUNCTIONALS OF A CONVEX SET 29

our directed index set), hence it converges to its supremum. This implies
that it is a Cauchy net, and since x,y in W implies that \\x-y\\^
*[/(*)—/O0]» say, we see that W itself is a Cauchy net in Z. Now, Z =
X n (K + z)\ since K is closed, Z must be complete and therefore W converges
to an element of y in Z. By continuity of / and of the norm, it is simple to
verify that y > x for all x in Wy i.e., W has an upper bound in Z. Zorn's
lemma then applies and our lemma is proved.
The ideas we have developed so far enable us to prove the density of
support points.
THEOREM 1. If C is a closed convex subset of a Banach space £ , then the
support points of C are dense in the boundary of C.
PROOF. If z is a point of the boundary of C and e > 0, choose y in E~C
Selected Papers of Errett Bishop Downloaded from www.worldscientific.com

such that 11 y — z \ \ < e/2 and choose (by applying the separation theorem to
any convex neighborhood of y which is disjoint from the closed set C) / in
by LA TROBE UNIVERSITY on 04/06/17. For personal use only.

E* such that | | / | | = 1 and sup/(C) ^f(y). By Lemma 1 there exists x0 in


C such that x0 e K(f, 2) + z and K(f, 2) -f x0 supports C at x0. Applying the
separation theorem to C and K + x0 shows that xQ is a support point of C.
Furthermore, since x0 — z e K and x0 e C, we have 11 xQ — z \ \ ^ 2[f(xQ) — f(z)] ^
2[f(y) ~ f(z)] ^ 2 \\y — 21| < e, which completes the proof.
To prove the density of support functionals we first prove that if k is large
and if g is of norm one and is non-negative on K(f,k) then / and g are
close together. This result is precisely stated in Lemma 3; the latter follows
from a known lemma [7] which is proved again here for the sake of complete­
ness.
LEMMA 2. Suppose that e > 0, | | / | | = l = ||£ll, and that \g(x)\ ^ e/2 when­
ever f(x) - 0 and \\x\\ S 1. Then either \\f + g\\ ^ e or \\f — g\\ ^ e.
PROOF. By the Hahn-Banach theorem we can choose h in E* such that
h=g onf"\0) and \\h\\ = snp \ g(U 0 f~\0)) |. Then, by hypothesis, \\h\\ ^
e/2. Furthermore, since g — h vanishes on / _1 (0) there exists a real number
a such that g — h — af. Hence \\g — af\\ = \\h\\ ^ e/2. Assuming a 2> 0, we
will show that \\f — g\\ S e. (Otherwise, the same proof applied to (—a)f
would show that | | / + £ | | g e.) If a ^ 1, then a"1 ^ 1 and
\\g-f\\ = II (1 - cTl)g + a~\g - af) || £ 1 - a'1 + a'1 \\g - af\\ .
Also, a.= \\af\\^\\g\\ + \\g - af\\ so
I - a" ^ (I + \\g - af\\)~l \\g - af\\ ^ \\g ~ af\\ .
1

Hence \\g-f\\ ^ 2 | | £ - a / | | g e. If 0 g a < 1, then


l l # - / l l ^ l l * - « / l l + lid - « ) / l l = l l * - « / l l + 1 - a
= lltf-«/ll + l l # I I H I « / l l £ 2 | | * - a / | | : S e ,
which completes the proof.
LEMMA 3. Suppose^ that 0 < e < 1, that | | / | | = 1 = | | # | | and that k > 1 + 2/e.
If g is non-negative on K(f,k)t then \\f — g\\ g e.

295
30 ERRETT BISHOP AND R. R. PHELPS

PROOF. Choose x in E such that | | x | | = l and f(x) > k~\l + 2/e), and
suppose that y in E is such that f(y) = 0 and \\y\\ g 2/e. Then ||*±3> li ^
1 + 2/e < */(#) = kf(x zty), so # zt y e K and hence g(x dzzy)^0. This implies
that \g(y) | ^ #(*) g ||* || = 1. Clearly, then, \g(y) | ^ e/2 whenever f(y) = 0
and \\y\\ ^ 1, so by Lemma 2, either | | / + # | | ^ e or | | / —#11 ^ £- Choose z
in £ such that || z || = 1 and f{z) > max (k"\ e). Then 26 if so £(2) ^ 0 and it
follows that | | / + g | | ^ (f + g)(z) > £> and our proof is complete.
We could now easily prove our density theorem; to include unbounded sets
C, it would be formulated somewhat as follows: If / is bounded on C, then
there is a support functional of C which is arbitrarily close to / . With little
care, however, we can do considerably more than this; the result is expressed
as follows: If / strictly separates the set C from a bounded set X, then there
is a support functional of C which is arbitrarily close to / and which strictly
Selected Papers of Errett Bishop Downloaded from www.worldscientific.com

separates X and C. More precisely:


by LA TROBE UNIVERSITY on 04/06/17. For personal use only.

THEOREM 2. Suppose that C and X are subsets of a Banach space E, that C


is closed and convex and that X is bounded and nonempty. If e > 0 and if f
in E*, 11/11 = 1, is such that sup/(C) < inif(X), then there exist g in E*,
\\g\\ = 1, and x0 in C such that | | / — # | | ^ e and g(x0) = sup#(C) < infg(X).
PROOF. Let r = sup/(C), 8 = inf f(X) and choose j9 such that r < £ < 8.
Consider the neighborhood V of X defined by X+ (8 - j9)J7= V. This is a
bounded set, and since inf/(£/) = - 1 , we have i n f / ( F ) = inff(X) - (8 - 0) = 0.
Let a = 1 + 2/e and choose 2 in C such that r - / ( * ) < (2«)"l(i3 - r). Let M
be larger than 2~1(j2 - r) and sup {I I .y - z\\:ye V} and let k = 2aM($ - r)" 1 .
(Note that k > a > 1.) Choose, by Lemma 1, a point x0 in C such that
K(f, k) + x0 supports C at xQ and x0 — 2 e if. We will show that V<z K + x0.
Indeed, if ye V, then | | y - * 0 || ^ llj> - *ll + IUo - 2II < M + \\x0 - z\\ S
M + kf(xQ -z)SM+k[r -f(z)] <M + k(2a)~lW - r) = 2M < 2aM = k(0 - 7) ^
kf(y — #0). By the separation theorem there exists g in E*,\\g\\ =h such
that sup g(C) = g(x0) ^ infg(K+Xo) ^ 'mig(V) = infg(X) - (8 - 0 ) < inf g(X).
Since 0 ^ inf g(K) and & > 1 + 2/e, it follows from Lemma 3 that | | / - g\\ ^ e.
A well known variant (and corollary) of the separation theorem substitutes
compactness for interior [5]: / / B and C are disjoint convex subsets of a locally
convex space, with B compact and C closed, then there exists a hyperplane which
strictly separates B and C, that is, there exists f in E* such that sup/(C) <
inf/(Z?). This result can be improved if E is a Banach space, giving the
following corollary of Theorem 2.
COROLLARY 1. If B and C are disjoint convex subsets of a Banach space E,
with B compact and C closed, there exist x in C and g in E* such that g(x) =
supg(C)< inf g(B).
An obvious corollary of the above variant of the separation theorem states
that a closed convex subset of a locally convex space E is the intersection of
all the closed half-spaces which contain it, that is, if x$C then there exists
/ in E* and a real number c such that the half-space H = {y:f{y) S c} con­
tains C but not x. We say that the half-space H supports C if C c H and

296
THE SUPPORT FUNCTIONALS OF A CONVEX SET 31

f(y) = c for some y in C. The proof of the following corollary is immediate


from Corollary 1.
COROLLARY 2. / / C is a closed convex subset of a Banach space, then C is
the intersection of all the closed half-spaces which support it,
COROLLARY 3, / / C is a closed convex subset of a separable Banach space E,
then C is the intersection of a countable number of its supporting closed half-
spaces.
PROOF. Let {**}£=.• i be a dense subset of E^C and let dn be the distance
from xn to C. By first applying the separation theorem to C and xn 4- dnU,
then applying Theorem 2 to C and xn + 2~ldnU, we can find gn in F * such
that gn supports C and sup£„(C) < inf gn(xn 4- 2~ldnU), for « = 1, 2, 3, •••.
Suppose, now, that xeE^C and let d be the distance from x to C. Choose
Selected Papers of Errett Bishop Downloaded from www.worldscientific.com

xn such that 11 x- xn\\ < 3"ld. Thenif j>€C, | | j > - *«|| ^ l l j > - * l l - | U - * n l l >
by LA TROBE UNIVERSITY on 04/06/17. For personal use only.

d - 3~ld = (2/3)d. Thus, dn ^ (2/3)d and hence \\ x - xn \\ < 2'ldn, which shows
that gn(x) > sup gn(C) and completes the proof.
If C is bounded, then every / in F * is bounded on C, so we obtain another
corollary to Theorem 2.
COROLLARY 4. / / C is a bounded closed convex subset of a Banach space Ef
then the support functionals of C are dense in E*.
If fe F*, we say that / attains its norm provided there exists x in U such
that ll/H = / 0 t ) . Since U is closed and convex, the proof of the following
corollary is immediate from the previous one. (This result was first proved
in [1].)
COROLLARY 5. / / E is a Banach space, then the set of f in E* which attain
their norm are dense in E*.
In all the above results we could have dropped the hypothesis that E be a
complete normed space provided we assumed that C itself be complete. This
follows from the fact that we applied Lemma 1 only to the set C. The follow­
ing result shows, however, that Corollary 4 (and hence Theorem 2) must fail
in an incomplete space.
THEOREM 3. If E is an incomplete normed linear space, then there exists a
bounded, closed convex subset C of E having nonempty interior such that the
support functionals of C are not dense in E*.
PROOF. Imbed £ a s a dense subspace of its completion F and identify (in
the obvious way) £ * and F*. Since E ^ F, there exists x in F~ E such that
| | x | | = 1. By applying the support theorem to x and the unit ball of F, we
can find / in F * such that | | / | | = 1 = / ( * ) . Let D = {yiyeF, \\y\\ £ 1 and
f(y) = 0} and let C in F be the convex hull of D and x, so that C is the
set of all elements of the form z = Xx + (1 — X)y, where yeD and Xe[0,1].
It is easily verified (using the compactness of [0,1]) that the convex set C is
closed; it also has nonempty interior. (If || z - (1/2)* || < 1/8, then 0 < f(z) < 1.

297
32 ERRETT BISHOP AND R. R. PHELPS

Define y by z = f(z)x + (1 — f(z))y; it is not difficult to verify that yeD so


that zeCf.) Let C = Cf n E; then C is convex, closed and has nonempty
interior relative to the (dense) subspace E. Clearly x&C; we will show that
if geE* and g(z) = sup g(C) for some z in C, then \\f-g\\^ 1/2. Write z =
Xx 4- (1 — ^)j as above; since ze C, we have z ^ A: SO that X < 1. Hence (since
sup#(C) = sup^(C)) it follows that g(x) ^ #(2) = (1 - k)g(y - x) + g(x) and
therefore g{y - x) ^ 0. Now 2 ^ | | * | | + ||;HI ^ II* - y\\ ^f(x- y) = 1 so
that lS(g-f)(y-x)^ | | # - / | | \\y-x\\ £2\\g-f\\. This shows that no
functional in E* which is within 1/2 of / can support C, and completes the
proof.
Note that Theorem 2 may also fail if we drop the assumption that the set
X be bounded. Indeed, if E is a separable nonreflexive Banach space, a
theorem of James [4] asserts the existence of / in £ * such that f(x) < 1 =
Selected Papers of Errett Bishop Downloaded from www.worldscientific.com

Il/lI for all x in U. Let X={y:f{y) ^ 2} and let C = U; then sup/(C) <
inf/GB), but if g separates C and B, then # must be some real multiple of/,
by LA TROBE UNIVERSITY on 04/06/17. For personal use only.

and hence cannot support C.


Another example (which is closely related to the preceding one) shows that
Theorem 2 may fail if we drop the hypothesis that / strictly separates X and
C. For, if £ is a separable nonreflexive Banach space, choose / i n E* as
above and let C = U and X = {*: 11 x 11 ^ 2 and /(*) ^ 1}. Then X and C are
disjoint bounded closed convex sets having nonempty interior, and sup/(C) =
1 = inff(X). Suppose there existed g in E* and x in C such that g(x) =
sup#(C) ^ 'mig(X). (Thus, g{x) = ||#||.) Let A be the convex hull of #and
X; since X has nonempty interior, so does A. Furthermore, it is not difficult
to see that 'mig(A) — g{x) = \\g\\. Since xeC we have /(#) < 1, and we can
choose z in C such that f(z) > 1 - 4_1[1 - /(*)]. Let ^ = 3_1(42 - x). Then 3>
is in the interior of X; indeed, ||y\\ ^ 3_1(4 | U | | + ||*||) = 5/3 < 2, while
/(>>) = (4/3)/(z) - (l/3)/(*) > (4/3) - (1/3)[1 - f(x)} - (l/3)/(*) = 1. Since xeA
and y is an interior point of At we conclude that (3/4) jy 4- (1/4)JC = 2 is an
interior point of A. But z is also in C, so there must exist a point M; in ^4
which is in the interior of C. This is impossible, since if we A then g(w) ^
\\g||, while #(*#) is less than | | # | | at the interior points w of C.
The first example given above shows that the conclusion to Theorem 2
fails if X is unbounded, even though it is assumed that C itself is bounded
and X is a linear variety (i.e., a translate of a linear subspace). If we assume
that X is a finite dimensional variety,- however, we get a valid result (which
is actually a corollary to Theorem 2). More' generally, we can assume that
X is a reflexive variety, that is, X = x -f M for some # in E and some sub-
space M of Is, where M i s a reflexive Banach space under the induced norm.
COROLLARY 6. Suppose that C is a bounded, closed convex subset of the Banach
space Ef that X is a reflexive linear variety in E, that e > 0 and that for
f in E*t 11/11 = 1, we have sup/(C) < inf f(X). Then there exists g in £*,
\\g\\ = 1, and Xo in C such that supg(C) = g(x0) < 'mfg(X) and \\f — g\\ < e.
PROOF. Since / is bounded below on X, it must be constant on X; writing
X — x + M as above, we then have / ( M ) = 0. Let El = E\M\ under the

298
THE SUPPORT FUNCTIONALS OF A CONVEX SET 33

usual factor-space norm, Ex is a Banach space and / can be regarded as an


element of norm one in E*. Let d be the image of C in Et under the
canonical mapping of E onto Eu and assume for the moment that d is closed
in Ei. Regarding X as a point in Eu we have s u p / ( d ) < f(X), so by Theorem
2 there exists # in E*f \\g\\ = 1, and XQ in d such that s u p ^ ( d ) = g(XQ) < g(X)
and | | / — g\\ < e. Since d is the image of C, there exists xQ in C such that
XQ = x0 + M. We can define g on £ in the obvious way; then \\g\\ = 1 and
| | / — g\\ < e in £*; since sup^(C) = g(xQ) < g(X), the corollary will be proved
once we have shown that d is closed. Suppose, now, that F in Et is in the
closure of d . Regarding F as a translate of M in £, this means that the
distance between Y and C in £ is zero. For each positive integer n, let G =
{x in £: d(x, C) ^ l/«}, where d(#, C) = inf {| | x — y \ | : j> e C}. The nested sets
G are nonempty, bounded, closed and convex, and the same is true of the
sets G fl Y. Since M is reflexive, the sets G 0 Y are weakly compact, and
Selected Papers of Errett Bishop Downloaded from www.worldscientific.com

therefore have nonempty intersection. Thus, there is a point y in Y such


by LA TROBE UNIVERSITY on 04/06/17. For personal use only.

that d(y,C) = 0, which implies that yeC and hence that F e d .


If we take E to be the Euclidean plane, C to be the convex region bounded
by a branch of a hyperbola, and X to be a translate (which misses C) of an
asymptote of C, we see that the above corollary is false if C is not assumed
to be bounded.
A special case of the above corollary could be formulated as follows: If
xu x2l ' — ,Xn are elements of the Banach space E, if e > 0, and if / in E*
vanishes on each %u then there exists a support functional g of the bounded
closed convex set C such that | | / — g\\ < e and g(xi) — 0, / = 1, 2, • • •, n.
(Simply let M be the linear span of the %i and let X — M + x for some x
such that sup/(C) </(#).) A dual result to this would be the following: If
fufzy - • >fn are elements of E*, if e > 0, and if z is an element of the boundary
of C such that fiz) = 0 for each i, then there exists a support point x0 of C
such that || 2 — x0 II < £ and /»(JC0) = 0, / = 1, 2, • • •, n. This dual result is valid
(even without the assumption that C be bounded) and is a generalization of
Theorem 1. In order to omit reference to the specific functionals fu ••-,/«,
we will work with the subspace N = {x: ft(x) = 0, i = 1, 2, - • •, n}; N is of finite
deficiency, and every subspace of finite deficiency has the same form as N.
We first prove a useful lemma.
LEMMA 4. Suppose that N is a closed subspace of finite deficiency in Ef that
C is a convex subset of E, and that XQ is a support point of C 0 N in the sub-
space N. Then xQ is a support point of C.
PROOF. Let n denote the deficiency of N; the proof will proceed by in­
duction on n. It is obviously true if n = 0; suppose, then, that n > 0 and
assume that the result is true for subspaces of deficiency n — 1. Let g be a
nontrivial functional in N* such that g(x0) = sup#(C (1 N). Note that if
C c N, then any functional which vanishes on N supports C at x0, so we can
assume that there exists an element y in C ~ N. Let A^; be the linear span
of N and y, and let C be the convex hull of C fl iV' and {x: xeN and g(x) ^
g(x0)}. Then C has nonempty interior relative to Nff C contains C fl N',

299
34 ERRETT BISHOP AND R. R. PHELPS

and x0 is in the boundary of C. By the support theorem, then, x0 is a


support point of C in Nr and therefore it is a support point of C fl N' in
N'. By the induction hypothesis, x0 must be a support point of C.
THEOREM 4. Suppose that N is a closed subspace of finite deficiency in the
Banach space E, that C is a closed convex set in E, that e > 0, and that z in
N is in the boundary of C. Then there exists a support point x0 of C such
that xQeN and \\z — xQ\\ < e.
PROOF. There are several cases to consider. First, if C is contained in a
proper closed subspace of £, then any functional which vanishes on this sub-
space supports C at each of its points, so z itself is a support point of C.
Assuming that C is not contained in a proper subspace, we consider whether
z is in the boundary (relative to N) of C fl N. If it is, then by Theorem 1
there is a point xQ in N which is a support point of C (\ N such that
Selected Papers of Errett Bishop Downloaded from www.worldscientific.com

\\z — x0 II < e. By Lemma 4, x0 is a support point of C in E. Finally, suppose


by LA TROBE UNIVERSITY on 04/06/17. For personal use only.

z is not in the boundary of C fl N; then there exists a neighborhood (relative


to N) of z which is contained in C 0 N. We will show that z itself is a
support point of C in E. There exists a point y in E such that the segment
[y, z[ = {x: x = Xy + (1 — ^)z, 0 < k g 1} is contained in £ ~ C; this follows
from the fact that N has finite deficiency and that z is in the boundary of C.
Let N' be the linear subspace spanned by N and y, and note that iV is a
hyperplane in Nf. We will show that N supports C fl N' at z, and hence (by
Lemma 4) z is a support point of C in £ . It suffices to show that the open
half-space {x + ry: xe N and r > 0} in N' is disjoint from C. Suppose that C
contained a point # + ry of this half-space. Since z is in the interior (relative
to N) of C fl iV, there would exist a point w of C and X in ]0,1[ such that
z = Ax -h (1 — ^)w. Hence the triangle with vertices z, M;, and x ■{■ ry would
be in C and (as can be easily shown) would also contain a point of [y, z[, a
contradiction.
The apparent duality between this theorem and Corollary 6 leads one to
conjecture that the theorem might still be true if it is merely assumed that
E/N is reflexive (rather than finite dimensional). The theorem fails, however,
under this weaker hypothesis; there is a well known example (see, e.g., [2, p.
160]) of a compact convex set C in a Hilbert space H and a line N'm H such
that C fl N = { 0 } , but 0 is not a support point of C (even though H/N is
reflexive).
All four of our lemmas have valid analogues in more general topological
vector spaces, although we do not know whether this is true of the theorems
themselves. (These questions will be the subject matter of another paper.)
For instance, it is unknown whether a closed convex subset of a complete
locally convex space must have any support points. (For those special classes
of closed convex sets C which are known to have support points, it is known
that the support points are dense in the boundary of C, e.g., if C has non­
empty interior, or if C is locally weakly compact [5].)

300
THE SUPPORT FUNCTIONALS OF A CONVEX SET 35

BIBLIOGRAPHY

1. E. Bishop and R. R. Phelps, A proof that every Banach space is subreflexive, Bull.
Amer. Math. Soc. 6 7 (1961), 97-98.
2. N. Bourbaki, Espaces vectoriels topologiqUes, Chapter V, Hermann, Paris, 1955.
3. N. Dimford and J. Schwartz, Linear operators, Part I, Interscience, New York,
1958.
4. R. C. James, Reflexivity and the supremum of linear functionals, Ann. of Math.
(2) 6 6 (1957), 159-169.
5. V. L. Kiee, Convex sets in linear spaces, Duke Math. J. 18 (1951), 443-466.
6. 1 Extremal structure of convex sets. II, Math. Z. 6 9 (1958), 90-104.
7. R. R. Phelps, A representation theorem for bounded convex sets, Proc. Amer. Math.
Soc. 1 1 (1960), 976-983.

UNIVERSITY OF CALIFORNIA, BERKELEY


Selected Papers of Errett Bishop Downloaded from www.worldscientific.com
by LA TROBE UNIVERSITY on 04/06/17. For personal use only.

301

Вам также может понравиться