0% found this document useful (0 votes)
139 views8 pages

Hahn-Banach Theorems

The document summarizes two theorems: 1) The Hahn-Banach extension theorem, which states that a linear functional on a subspace can be extended to the whole space while maintaining certain properties. The proof uses a claim about extending to finite-dimensional subspaces and then taking a direct limit. 2) The Hahn-Banach separation theorem, which provides a way to separate two disjoint convex sets using a linear functional. The proof constructs a sublinear functional and then extends it using another application of Hahn-Banach extension.

Uploaded by

Dragan Rukavina
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
139 views8 pages

Hahn-Banach Theorems

The document summarizes two theorems: 1) The Hahn-Banach extension theorem, which states that a linear functional on a subspace can be extended to the whole space while maintaining certain properties. The proof uses a claim about extending to finite-dimensional subspaces and then taking a direct limit. 2) The Hahn-Banach separation theorem, which provides a way to separate two disjoint convex sets using a linear functional. The proof constructs a sublinear functional and then extends it using another application of Hahn-Banach extension.

Uploaded by

Dragan Rukavina
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 8

Talk by Andreas Naes Aaserud for the logic seminar (290D) on 1 May 2012.

Hahn-Banach theorems

Theorem. (Hahn-Banach extension) Let X be a subspace of a real linear space Y and let p : Y
R be a sublinear functional. Then for any linear functional f : X R satisfying f (x) p(x) for
all x X, there is a linear functional f: Y R such that f|X = f and f(y) p(y) for all
y Y.
Proof. We will need the following
Claim. Let a Y \ X be given. Then f extends to f: span(X {a}) R such that f(y) p(y)
for all y span(X {a}).
We refer to Folland, p. 158, for the standard proof of this fact.

Next, consider Wa = {Z FDX (Y ) : a Z} for a Y . (Here, FDX (Y ) is the set of subspaces


of Y that are finite-dimensional over X, i.e., are spanned by X and some finite subset of Y \ X.)
As {Wa }aY has FIP, there exists Z FDX (Y ) such that Y Z Y .
The claim implies that, for any Z0 FDX (Y ), there exists f: Z0 R such that f|X = f
and f(z) p(z) for all z Z0 . By transfer, there exists f: Z R such that f|X = f and
f(z) p(z) for all z Z.
Thus f = stf|Y is the desired extension of f .

Theorem. (Hahn-Banach separation) Let X be a real normed space, A, B X non-empty disjoint


convex sets, where int(A) 6= . Then there exists f X 0 and R such that f (A) (, ]
and f (B) [, ). If A is open, we may require that f (A) (, ).
Proof. Put C = A B. Then C is convex with non-empty interior and 0
/ C. Fix c int(C)
and put K = C c.
Define p : X [0, ) by p(x) = inf{t R+ : x tK}.
Claim. p is a sublinear functional.
p well-defined: Given x X, x/n 0 as n . As 0 int(K), x/n K for some n. Thus
x nK for some n.
p positive homogeneous: Let [0, ) be given. If = 0 then p(x) = 0 = p(x). If > 0
then
p(x) = inf{t/ R+ : x (t/)K} = p(x).
p subadditive: If x sK and y tK then x + y = (s + t)(s/(s + t)x + t/(s + t)y) (s + t)K.
Thus p(x + y) s + t. It follows that p(x + y) t p(x) whenever y tK, hence that
p(x + y) p(x) + p(y).

Define a linear functional f0 : span{c} R by f0 (rc) = rp(c).


Claim. f0 (x) p(x) for all x span{c}.
For r < 0, we have that f0 (rc) = rp(c) = p(rc), as p is positively homogeneous. For s > 0, we
claim that f0 (sc) p(sc). Indeed, 0 = p(c + c) p(c) + p(c) so that sp(c) (s)(p(c)) =
sp(c). Thus f0 (sc) = sp(c) sp(c) = p(sc).

By Hahn-Banach extension, we may extend f0 to f : X R with f (x) p(x) for all x X.


As p(x) = inf{t : x tK} 1 whenever x K, it follows that f (x) 1 for all x K. As
0 int(K), there exists  > 0 such that B(0, ) K. Hence kxk  |f (x)| 1. It follows that
kxk /n |f (x)| 1/n, so if x 0 then f (x) 0. Hence f is continuous.
1

Note moreover that f (c) = f (c) = f0 (c) = p(c) 1. Indeed, c


/ K and tK K for all
t (0, 1), because 0 K and K is convex.
Finally, if x A and y B then x y c K so that
f (x) f (y) + 1 f (x) f (y) f (c) = f (x y c) 1,
i.e., f (x) f (y). Thus = inf{f (y) : y B} works. The statement concerning open A follows
from the Open Mapping Theorem.

Corollary 1. Let X be a real normed space, C X convex, and a X with dist(a, C) > 0.
Then there exists f X 0 such that f (a) < inf{f (x) : x C}.
Proof. Take B = C and A = B(a, ) such that A B = .

Corollary 2. Let X be a real normed space, Y X a closed subspace with a Y and dist(a, Y ) >
0. Then there exists f X 0 such that f (a) = 1 and f (Y ) = {0}.
Proof. Take B = Y and A = B(a, ) such that A B = . Then there exists f X 0 and R
such that f (y) for all y Y while f (a) < . As Y is a subspace, we must have f (y) = 0 for
all y Y . Finally, replace f by f /f (a). (Note: f (a) 6= 0 as otherwise > 0 so f (0) > 0.)

General non-standard hull

Setup. Let X be an internal linear space over R and let W be a family of internal seminorms
on X such that |W | < . Define an equivalence relation on X by
x W y p W : p(x y) 0.
We define the monad with respect to W (of x X) by
W (x) = {y X : x W y}
and the finite part of X with respect to W by
FinW (X) = {x X : sup st(p(x)) < }.
pW

b W = {W (x) : x FinW (X)} and


We define the non-standard hull of X with respect to W by X
W
b
define k kW : X R by
k
xkW = sup st(p(x)),
pW

where x
= W (x).
Remarks. (i) For x X, W (x) = x + W (0). If x FinW (X), then W (x) = {y FinW (X) :
x W y}.
(ii) W (0) is an R-linear subspace of FinW (X).
b W , k kW ) is a Banach space.
Lemma. (X
b W obtains a vector space structure from the identification X
b W = FinW (X)/W (0).
Proof. (i) X
(ii) k kW is a norm, as
k
xkW = 0 p W : st(p(x)) = 0 p W : p(x) 0 x W (0) x
=0
and
k
xkW = kxk
c W = sup st(p(x)) = sup st(||p(x)) = sup ||st(p(x)) = ||k
xkW .
pW

pW

pW

The triangle inequality is clear.


(iii) k kW is complete: Let {c
an }
n=1 be a Cauchy sequence. Given m N choose km such
that
n > km kc
an ad
km kW < 1/(2m).
Extend {c
an } to {c
an }n<N , N N \ N.
For p W and m N, define
Fp,m = {an : n < N p(an akm ) < 1/(2m)}.
Claim 1. {Fp,m }pW,mN has FIP. (Given Fp1 ,m1 , . . . , Fpr ,mr , take an with n > maxj=1,...,r kmj .)
So by saturation, there exists a p,m Fp,m .
Claim 2. n, m N : (n > km ) suppW st(p(a an )) 1/m.
Let n, m N be given. Suppose n > km ; let p W . Then
st(p(aan )) st(p(aakm )+st(p(akm an )) < 1/(2m)+kc
an ad
km kW < 1/(2m)+1/(2m) = 1/m.
Fixing n > k1 in claim 2 shows that a an FinW (X) so that a FinW (X). Moreover, claim 2
shows that k
a ac

n kW 0 for n .
b W = FinW (X)/W (0) as Banach
Proposition. W (0) is closed in (FinW (X), kb kW ) and X
spaces.
Proof. Trivial. See also Proposition 2.6 concerning norm non-standard hulls.

Weak non-standard hull

Let X be a real normed space. We will apply the notions from the previous section to X .
Consider the weak topology on X generated by the semi-norms pf (x) = |f (x)| for f X 0 . For this
family W of seminorms,
x W y f X 0 : f (x) f (y).
Proposition. Let C X be a set. Then
(i) C weakly closed C is norm-closed;
(ii) C is norm-closed and convex C is weakly closed.
Proof. (i) Trivial.
(ii) Let c C , X 3 a W c. We want to show that a C. (By Proposition 1.14(ii), this will
suffice.) Suppose a
/ C. By the Hahn-Banach separation theorem (cf. corollary 1), there exists
f X 0 such that f (a) < inf xC f (x). By transfer, f (a) < inf xC f (x). In particular, f (a) < f (c)
so that f (a) 6 f (c). As a W c, this is a contradiction.

X 0 } and call X
c
Setup. We put now W = {pf : f B

the weak non-standard hull of X.

Define an infinite-valued seminorm stk kX : X [0, ] by setting stkxkX = if kxkX


/
Fin(R ). On the quotient space X /W (0), we define the quotient seminorm
kx + W (0)kq = inf{stkx + ykX : y W (0)} [0, ]
for x X . The finite part of X /W (0) is defined as
Finq (X /W (0)) = {x + W (0) : x X kx + W (0)kq < }
3

and forms a subspace of X /W (0).


We will use

Proposition. Let X be a normed space and f X 0 . Then there exists a BX


such that

kf k f (a) [0, ) .

Proof. Consider the internal sets En = {x BX


: kf k |f (x)| + 1/n}, n N. As {En }nN has

the FIP, there exists a0 BX such that kf k |f (a0 )|. As transfer yields x BX
: |f (x)| kf k,

it follows that kf k |f (a0 )|. Choose next F such that f (a0 ) = |f (a0 )| and put a = a0 .

Then a BX
does the trick.


Theorem. We have that


(i) x FinW (X ) : k
xkW = kx + W (0)kq ;
c
(ii) X

= Finq (X /W (0)) as Banach spaces.

Proof. (i) Fix a FinW (X ). Then, for y W (0), we have that


k
akW = sup st|f (a)| = sup st|f (a + y)| stka + ykX .
0
f B
X

0
f B
X

Hence k
akW inf yW (0) stka + ykX = ka + W (0)kq .
We next prove the reverse inequality. Without loss of generality, we may assume that k
akW = 1.
Assuming this, we wish to show that ka + W (0)kq 1.
Write, for f X 0 , rf = st(f (a)).
Claim. For A SX 0 finite and n, m N, there exists c (1 + 1/m)BX such that
f A : |f (c) rf | 1/n.
Suppose for contradiction that the claim fails for A, n, m. Then
^
x X :
(|f (x) rf | 1/n) x
/ (1 + 1/m)BX .
f A

= {x X : f A |f (x)rf | 1/n}. Then A


B
= by the previous
Put A = (1+1/m)BX and B
are convex and A is open. Hence the Hahn-Banach separation theorem
line. Moreover, A and B
(, ) and f(B)
[, ).
implies that there exist f X 0 and R such that f(A)
Hence, for any x, y X, if kxk < 1 and f A [|f (y) rf | 1/n], we have that
f(x) <

< f(y).
1 + 1/m

()

By rescaling f if necessary, we may assume that kfkX 0 = 1.

The previous proposition implies that there exists x BX


such that 1 = kfkX 0 f(x) R .
Apply transfer to () to get that

1 f(x) <

< f(a) |f(a)|.


1 + 1/m

Hence k
akW = supf BX 0 st|f (a)| > > 1, a contradiction.

Put Ef,n,m = {x X : x (1 + 1/m)BX


|f (x) rf | 1/n} for f SX 0 and n, m
N. The claim implies that {Ef,n,m }f,n,m has FIP. Thus we get by saturation that there exists

c f,n,m Ef,n,m . This c satisfies kck 1 and f (x) rf = stf (a) f (a) for all f X 0 , i.e.,
(c a) W (0). It follows that
ka + W (0)kq =

inf
yW (0)

stka + ykX stka + (c a)kX = stkck 1,

showing what we want.


W

c Finq (X /W (0)) as sets and, by part (i), also as Banach spaces. [In
(ii) Note that X
other words, the map x
x + W (0) is a linear isometric embedding.] To see that we have
equality [surjectivity], let x X with kx + W (0)kq < be given. Choose y W (0) such that
X 0 , we have that
stkx + yk < . Then, for any f B
st|f (x)| = st|f (x + y) f (y)| = st|f (x + y)| stkx + yk < .
Hence x FinW (X ), showing what we want.
c
Corollary. For all x
X

there exists y W (0) such that k


xkW kx + ykX .

Proof. Apply saturation to the sets Ep,n,m = {y X : > kx + ykX 1/n p(y) < 1/m}
for p W and n, m N, where := k
xkW .

c
Proposition. The map X X

given by x 7 x
is an isometric embedding.

Proof. Note that k


xkW kx + ykX kxk (where y W (0)). By Hahn-Banach extension, we
may extend 0 : span{x} R given by 0 (rx) = rkxk to a linear functional : X R of norm
1 and with (x) = 0 (x) = kxk. Thus
k
xkW = sup st|f (x)| st|(x)| = kxk,
0
f B
X

proving what we want.

For the proof of the following theorem, we will need


Theorem. (Helly) Let X be a normed space, let I be a finite set, and let {fi }iI X 0 and
(i )iI F. Then, given r R+ , TFAE:
(i)  R+ x X [(kxk r + ) (i I : fi (x) = i )];
P
P
(ii) {i }iI F : | iI i i | rk iI i fi k.
Corollary. Let X be a normed space, let I be a set, and let {fi }iI X 0 and (i )iI F. Then,
given r R+ , (ii) imples (i), where
(i) x X [(stkxk r) (i I : fi (x) i )];
P
P
(ii) J P fin (I){i }iJ F : | iJ i i | rk iJ i fi k.
Proof. By Hellys Theorem, there exists, for each n N and each J P fin (I), an element x X
with (kxk r + 1/n) (i J : fi (x) = i ). Thus it suffices to apply saturation to the family
En,J = {x X : (kxk r + 1/n) (i J : fi (x) = i )} of internal sets.

Theorem. Let X be a normed space. Then the canonical embedding 0 : X X 00 extends to an
W
c X 00 .
isometric isomorphism : X
W

c X 00 by (
Proof. Define : X
x) = st(x). Then is a well-defined injective linear
functional. Clearly, |X = 0 .
Let F X 00 be given. Note that, for each finite set J X 0 and each { }J R, we have
that






X

X






F () kF kX 00

.

J

J

5

Hence the corollary to Hellys theorem implies that there exists a X such that
stkak kF kX 00 X 0 : (a) F ().
In particular, a Fin(X ) FinW (X ) and, for each X 0 , (
a) = st(a) = F (), i.e.,
(
a) = F . Moreover,
k
akW = sup st|(a)| = sup |(
a)| = sup |F ()| = kF kX 00 ,
0
B
X

0
B
X

0
B
X

showing what we want.

Theorem. Let X be a normed space and let K, C X be disjoint non-empty sets such that K
is weakly compact and C is norm-closed and convex. Then dist(K, C) > 0.
Proof. Suppose for contradiction that dist(K, C) = 0. Then, for any n N, there exist a K
and c C such that ka ck < 1/n. Apply saturation to the sets En = {(a, c) K C :
ka ck < 1/n} to find a K and c C such that a c. As K is weakly compact, there exists
b K such that b W a. As b
/ C, there exists r > 0 such that A = B(b, r) is disjoint from C.
The Hahn-Banach separation theorem implies that there exist f X 0 and R such that
f (b) < x C : f (x) .
By transfer, f (c) . In particular, f (b) 6 f (c).
On the other hand, b W a implies that f (b) f (a) so that f (a) 6 f (c), contradicting the fact
that a c.


Weak* topology

X . For
Consider the weak* topology on X 0 generated by the semi-norms pa () = |(a)| for a B
this family W of seminorms,
W x X : (x) (x).
Note for the next proof that (X )0 = (X 0 ) , where (X )0 is the space of internally continuous
linear functionals X R .
X } of
Proposition. Let X be a normed space and equip (X 0 ) with the family W = {pa : a B
W
0 )
\
seminorms. Then the map : X 0 (X
given by 7 is an isometric isomorphism.
Proof. Note that is linear and that
k()k = kk
= sup st|(a)| = kkX 0 .
X
aB

0 )
\
Let next (X
be given and put = st |X. Then is a linear functional on X and
for all x B
X so that is also continuous. Finally, (x) (x) for each
|(x)| = st|(x)| kk
x X. Hence ( ) W (0) and = = ().


X 0 is weak*ly compact.
Theorem. (Alaoglu) Let X be a normed space. Then B
c (under x 7 x
Proof. We will repeatedly use the fact that R
). Moreover, we use the usual
=R
non-standard characterization of compactness.
X 0 ) = B
(X 0 ) be given. We must show that there exists B
X 0 such that W
Let (B
.
6

)0 (X
\
[
c )0 defined by ()
Consider the map : (X
x = (x).

Claim. is a well-defined linear isometric isomorphism.


i.e., . Then (x) (y) for all x, y Fin(X ) with x y, so ()
Suppose = ,

[
is uniquely determined. Moreover, |(x)| |(x)| and k
xkXc kxkX for all x X . Hence
k()k
= kk
so that is a well-defined linear isometry.
c )0 be given. Put = |X under the usual
It remains to show that is onto. Indeed, let (X
identification. Then the non-standard extension of , which we will denote by the same symbol
)0 . Finally, as (x) = (x) for all x X, transfer
\
as usual, belongs to Fin((X )0 ). Hence (X

[ (x) = (x). As (x) and (x)


[ both
implies that (x) = (x) for all x X . Thus (x)
belong to R, we get that ()
= , showing surjectivity.

d R

c = R.
X 0 . Given x X, put = (x)
Under this identification, put = |X.
Then B
x) = (x) under our identifications. Thus W and we are done. 
Then (x) = (

Riesz Representation Theorem

In
ba(N)
= (` )0 given by 7 , where (f ) =
R the followingwe will use the identification
0
f d for f ` (N). By transfer, each ((` ) ) is of the form for some ba(N) .
N
Note also that we are here working with complex scalars, unlike in previous parts of the talk.
Moreover, we use without proof the following theorem on general Loeb measures:
Theorem. Let (, B, ) be an internal measure space, where is a finite internal complex measure
and B need only be an algebra of sets. Then st extends uniquely to a -additive measure on
B, where B is the -algebra generated by B.
Theorem. Let be a compact topological space. Then the map : M() C()0 given
by
R
7 , where M() is the space of complex regular Borel measures on and (f ) = f d,
is a linear isometric isomorphism.
Proof. We need only show that is onto so fix C()0 .
Step 1. We claim that there exists a hyperfinite set H such that H and st(H) = .
Consider, for , the set W = {H P fin () : H}. As {W } has FIP, there exists
a hyperfinite set H such that H . As is compact, st = ; hence st(H) = .
Step 2. We identify internally H with [0, N ] N for some N N . Thus we may embed
` (H) ` (N) internally. For f C(), we put f = f |H ` (H) ` (N) . For
= span(f1 , . . . , fn ) ` (N) . For f C(),
X FD(C()) with basis {f1 , . . . , fn }, put X

x H, we have f (x) = f (x) f (st(x)) so that stf(x) = f (st(x)). Moreover,


{f1 , . . . , fn } C() lin. indep. {f1 , . . . , fn } ` (N) lin. indep.
by transfer.
Step 3. If X = span{f1 , . . . , fn } C(), with f1 , . . . , fn lin. indep., then we define an internal
C by fi 7 (fi ). This functional is independent of the choice of basis.
linear functional X : X
By Hahn-Banach extension, X extends to X (` (N) )0 with kX k = kX k. Note that, for
f X, X (f) = X (f) = (f ).
Step 4. Apply saturation to the family of sets EX,n = { (` (N) )0 : |kk kk| < 1/n f
X : (f) = (f )} for X FD(C()) and n N to find (` (N) )0 with kk kk and
(f) = (f ) for all f C().

Step 5. By the remark before the proof, we may choose ba(N) such that = . Let be
the restriction of to the internal subsets of H. As
||() ||(N ) = kk = kk kk R,
is a finite internal complex measure. Thus st extends to a unique -additive measure L().
Put = L() st1 , defined on the Borel subsets of . Then M() and, for each f C(),
we get
Z
Z
Z
Z
Z

(f ) = (f ) =
f d =
f d =
st f dL() =
f st dL() =
f d.
N

This completes the proof.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy