Download as pdf or txt
Download as pdf or txt
You are on page 1of 64

Applications of Hahn

Banach Theorem
Modern Approach
A.C.Pandey

1
E: normed vector space, assumed
to be real for definitions
Known: For each x  E , y   E  s.t.
y   x and y , x  x
2

y
Taking x  We have
x
For each x  E , x  E  s.t.
x  1 and x, x  x
2
Corollary 1
If x  y  E , then x  E  s.t.
x, x  x, y

E separates po int s of E
Proof:
x  E  s.t ,
x, x  y  x  y  0
 x, x  y  0
 x, x  x, y  0
 x, x  x, y 3
Corollary 2 Proof in next
page
E can be embeded isometrically in E.

This corollary implies that


We may consider E as embedded in

E 
 as normed space, then is a E
complete space which is the
completion of E.
4
For x  E , define j ( x ) in E  by
j ( x ), x   x , x for x  E 
 j : E  E  is linear and j ( x )  E  .
If x1  x2 , by Corollary 1
 x   E  s.t. x , x1  x , x2
 j ( x1 ), x   j ( x2 ), x 
 j ( x1 )  j ( x2 )
Hence j is one to one.
Claim : j is an isometry; i.e. , j( x)  x x  E
Pf : For xE
j ( x )  sup j ( x ), x   sup x , x  sup x  x  x
x  1 x  1 x  1

On the other hand ,


x   E  s.t. x   1 and x , x  x
 x  x , x  j ( x ), x   j ( x ) x   j ( x )
5
Hence j( x)  x
A dual variational principle
Let F be a v.s.s of E, and let

F  x  E  x, x  0  x  F


Then for any x  E , we have
dist( x, F )  inf x  y
yF

 max x, x  max x, x


xF  , x 1 xF  , x 1 6

For any x  F with x  1, we have
for y  F
x, x  y  x  y x  x  y
 x  y  x, x  y
 x  y  x, x
Hence inf x  y  sup x, x
yF x F  , x  1

On the other hand ,


define on x  F  x  y   R, y  F   M
a linear functional f (x  y )   dist( x,7F )
 dis( x, F )
f  sup
M
R , yF x  y
dis( x, F )
 sup 1
R , yF y
x

Hahn  Banach Theorem implies that
f can be extended to be defined on E
with the same norm, still denoted by f .
i.e. f  E  and f E
1

But f F and f , x  dis( x, F )
Hence dis( x, F )  max x , x . 8
x F , x  1

Example
Let E  C0,1
Why?
the space of all continuous functions on E
See next
For x  E , x  max x(t ) page
t0,1

Define f  E by  f  1
1
E
1
f , x  02 x(t )dt   x(t )dt
1 Actually, f E
1
2

1
f , x  0 x (t ) dt  x
9
For   0, Let x be the following function.

1 1
  1
2 2

-1

10
x 1

and
f ( x )  1    lim f ( x )  1
  0 

 f E
 sup f ( x )  1
x 1

Hence f E
1

11
Claim
In this example

There is no x  E with x  1 s.t.


f E
 f ,x 1

12
Exercise
Suppose that f  E  does not have
the property that x  E with x 1
s.t f  f ,x
Show that if we let
F  x  E f , x  0 ker f
then for any x  E , x  F ,
there is no y  F s.t
dis( x, F )  x  y 13
x  y that
Suppose x  E , x  F , y  F such that
 dis( x , F )
dis , F )  x x,yx , then
 ( xmax
x F  , x  1

f  f
 ,x , sin ce f F , 1
f f
1
 f,x
f
 f,x  f  x y (1)
On the other hand ,
f,x  f ,x y , sin ce y  F  ker f
 f  x y
14
Hence f ,x  f x y (*)
x y
Let z  , then
x y
zE, z  1, and
x y
f ,z  f,
x y
1
 f ,x  y
x y
1
 f , x , sin ce y  F  ker f
x y
1
 f x y , by (*)
x y
 f ,
15
which is impossible
Applications of Mazm-
Orlich Theorem

16
Pf (S )
is the space of the probability
measure of S
s1 ,, sk
p
1 ,, k
 p( si )  i ,  i  1,, k
k 1
where   (1,, k )  
17
 :S E v. s

s1 ,, sk
p  Pf ( S )
1 ,, k
k
 ( p )  s  dp   i ( si )
i 1

18
Mazm-Orlich Theorem
pi : Ei  R sublinear i  1,2
 i : S  Ei i  1.2
( I )li  Ei *
with li  pi  i  1,2
s.t l1 ( 1 ( s))  l2 ( 2 ( s ))
 ( II ) For p  Pf ( S )
 ( p1 (  1 ( p ))  p2 ( 2 ( p )) 19
Mazur-Orlicz (1953)
S : arbitary set E : real vector space
 :S R  :S E
q: E  R sublinear

(I)   E *
with   q such that
 ( s)   ( s)  s  S

 ( II )  ( p )  q ( p )   p  Pf ( S ) 20
Take E1  R, E2  E , q1 (t )  t , q2  q, 1   ,
and  2  
Claim : 1  E and 1 ( x )  q1 ( x )  x  1 ( x )  x
*

pf :
1 ( x )  x  x  E
 1 (  x )   x  x  E
  1 ( x )   x  x  E
 1 ( x )  x  x  E
Hence 1 ( x )  x  x  E 21
Corollary 1
Let E ,q, be as in Mazur-Orlicz Theorem

then

sup inf ( ( s))  inf q( ( p))


 q sS pPf

22
Let   R and define  ( s )    s  S
By Mazur  Orlich Theorem
(*) ( I )    E with   q s.t.
*

  ( ( s ))  s  S
 ( II )   q( ( p ))  p  Pf ( S )
Take   inf q( ( p ))
pPf ( S )

then ( II ) holds and implies


  E *
with   q s.t. inf q( ( p ))  ( ( s )) sS
pPf ( S )

then inf q( ( p ))  sup inf ( ( s ))


pPf ( S )   q sS
23
Given   0   E with   q s.t.
*

sup inf ( ( s ))    ( ( s ))  s  S


  q sS

By (*) sup inf ( ( s ))    q( ( p ))  p  Pf ( S )


  q sS

sup inf ( ( s ))    inf q( ( p ))


  q sS pPf ( S )

By taking   0, we have
sup inf ( ( s ))  inf q( ( p ))
  q sS pPf ( S )

Hence sup inf ( ( s ))  inf q( ( p ))


  q sS pPf ( S )
24
Corollary 2
If in Corollary 1  sastisfies the condition:
For each p  Pf ( S ),   0
there is sS such that

( ( s ))  ( ( p ))     E *
with   q

then
sup inf ( ( s))  inf q( ( s))
 q sS sS
25
Since sup inf ( ( s ))  inf q( ( p ))  inf q( ( s )) ,
  q sS pPf ( S ) sS

it is sufficent to show that inf q( ( s ))  inf q( ( p ))


sS pPf ( S )

 inf q( ( p )) if inf q( ( p )) is finite


 pP ( S ) pP ( S )
Let    f f

arbitary if inf q( ( p ))  


 pP ( S ) f

Given   0,  p  Pf ( S ) s.t. q( ( p ))    


For this  and p ,  s  S s.t.
( ( s ))  ( ( p ))      E * with   q
For this s, Hahn  Banach Thm implies that
 ˆ  E * with ˆ  q s.t. ˆ( ( s ))  q( ( s ))
then q( ( s ))  ˆ( ( s ))  ˆ( ( p ))    q( ( p ))      2
then inf q( ( s ))    2
sS
By taking   0, we have inf q( ( s ))  
sS
i.e. inf q( ( s ))  inf q( ( p )) 26
sS pPf ( S )
Example p.1
S: arbitary set f1,, f n : S  R
E  R , q( x )  max xi ,  x  ( x1 ,, xn )
n
1i  n

 : S  E defined by
 ( s)  ( f1( s),, f n ( s))  s  S Then
n
sup inf   i f i ( s )  inf max f i ( p )
 n 1 sS i 1 pPf ( S ) 1i  n 27
n 1
Since E  R , for any   E    
n
s.t.
n
( x )    i xi  x  ( x1 ,, xn )  R n

i 1
n
then ( ( s ))    i f i ( s )  s  S , and
i 1
q( ( p ))  max f i ( p )
1i  n
By Corollary 1
sup inf ( ( s ))  inf q( ( p ))
  q sS pPf ( S )

n
 sup inf   i f i ( s )  inf max f i ( p ) 28
n 1 sS i 1 pPf ( S ) 1i  n
Example p.2
 is q-convex

 p  Pf ( S ), s  S s.t.
( ( s))  ( ( p))   E,   q
n n
n 1
 i f i ( s )  i f i ( p )   
i 1 i 1

 fi ( s)  fi ( p) i  1,, n 29
Example p.3
In particular, S is a convex set in a linear map

fi : S  R is convex i.e.
 r  r
f i    j s j    j i j
 f ( s )  s1 ,  , s r   E ,   n 1
, i  1,, n
 j 1  j 1

Then
n n
sup inf i f i ( s )  inf max i f i ( s )
 n 1 sS i 1 sS n 1 i 1

This implies von Neumann Minimax Theorem


30
s1 ,, sr
For any p   Pf ( S ), i  1,, n
1 ,, r
r  r 
f i ( p )    j f i ( s j )  f i    j s j 
j 1  j 1 
r
Let s    j s j , then f i ( p )  f i ( s )
j 1

then  is q  convex
then by Corollary 2, we have
sup inf ( ( s ))  inf q( ( s ))
  q sS sS

n n
 sup inf  i f i ( s )  inf max f i ( s )  inf max  i f31i ( s )
 n 1 sS i 1 sS 1i  n sS n 1 i 1
To show that von Noumann Mini max Theorem
Given aij nm ,
Let S  m 1 , E  n 1 , and for i  1,, n
Define f i : S  R by
m
fi (  )   aij  j   m 1
j 1

Then f i is convex
By previous result, we have
n n
max min
 n 1
  m 1
  i f i (  )  min min   i f i (  )
   m 1 n 1
i 1 i 1
n m n m
 max min
 n 1
  m 1
 i  aij  j  min min   i  aij  j
  m 1 n 1
i 1 j 1 i 1 j 1
n m n m
 max min
 n 1
  m 1
  aij i  j  min max   aij i  j 32
   m 1 n 1
i 1 j 1 i 1 j 1
Duality map p.1
J(x) is
Let E be a real reflexive Banach w-compact
space. For

xE
(see next page)


J ( x)  y  E  : y  x ,  y, x  x
2

J is a Duality map. If E is a Hilbert space, then

J ( x)  x
33
Since E is reflexive and J ( x ) is bounded , convex
and closed , J ( x ) is w  compact
 To show that J ( x ) is convex. 
 
 For y1 , y2  J ( x ) and (1 , 2 )   
 

 1 1y   y
2 2   y
1 1   y
2 2  1 x   2 x  x 
 ( y   y , x )   ( y , x )   ( y , x )   x 2   x 2  x 2 
 1 1 2 2 1 1 2 2 1 2 
  x 2  ( y   y , x )   y   y  x 
 1 1 2 2 1 1 2 2 
  x  1 y1   2 y2 
 
 Hence 1 y1   2 y2  J ( x ) 

34
Lemma p.1

Let S be a compact convex subset of a


linear function
topological linear space. Define+constant
q : C ( S )  R by q( f )  sup f ( s )
sS

If   C ( S ),   q
and F is the space of all affine functions then

y0  S s.t.  ( f )  f ( y0 ) f  F 35
For any f  F , let A f  y  S ; ( f )  f ( y )
Claim : A f  
 ( f )  q( f )  max f ( s ) 
 sS 
   ( f )   (  f )  max  f ( s )   min f ( s )
 sS sS 
 min f ( s )   ( f )  max f ( s ) 
 sS sS 
  y  S s.t. f ( y )  min f ( s )   ( f ) 
 sS 
 Hence A f   
Now to show that  A f  
f F

It is sufficient to show that



 Af  
i
for every finite subset  f1,, f    F 36
Given any finite subset  f1,, f   of F
let hi  f i   ( f i ) for i  1,, 
By Corollary 1

min max hi ( p )  max min   j h j ( s )
pPf ( S ) 1i    1 sS j 1

 min max  f i ( p )   ( f i )
pPf ( S ) 1i  

 max min   j  f i ( s )   ( f i )
 1 sS j 1

  
 max min    j f i ( s )   (   j f i )  0
 sS j 1
 1

 j 1 
then  s0  S s.t. max  f i ( s0 )   ( f i )  0
1i  

Hence s0   A f i
37
i 1
Theorem p.1

Let E be a real reflexive Banach space

a : E  E  R is bilinear such that

(i) There is c>0 such that

a ( x, y )  c x  y x  E , y  E 
(ii)   0 s.t. min a( x, y )   , x  E , x  1
yJ ( x )
38
Theorem p.2
Then for each   E there is a unique


y0  E with y0  such that

( x)  a( x, y0 ) x  E
39
May assume that   0
Let K be the closed unit ball in E ,
then K is w  compact , sin ce E is reflexive
Let G  a ( x,)  ( x ); x  E
Every function in G is w  continuous and every
finite linear combination of elements in G is still

in G. Let S  K , where  

then S is w  compact subset of E
Define q : C ( S )  R by q( f )  max f ( s ) f  C ( S )
sS
Let  : G  C ( S ) by  ( g )  g g  G
40
Claim : q( ( p ))  0 p  Pf (G )

g1 ,, g r 
For any p   Pf (G ) 
1 ,, r 

 r
 r

q( ( p ))  q  i gi   max  i gi ( s )
 i 1  sS i 1 
 max a ( x0 , y )  ( x0 ) for some x0 
yS  K 

  max a ( x0 , y )  ( x0 ) 
yK

 
 max a ( x0 , y )   x0
x0 y  x 

0

 
 max a ( x0 , y )   x0
x0 yJ ( x ) 

0


 x0   x0  x0      0 
2


41
x0 
q( ( p ))  0 p  Pf (G )
by Mazur Olicz Thm
   C ( S ) with   q s.t.
 ( ( g ))  0 g  G
  ( g )  0 g  G
G is linear
  ( g )  0 g  G
By Lemma y0  S s.t.  ( f )  f ( y0 ) f  F
where F is the space of all affine functions

then y0  and 0   ( g )  g ( y0 ) g  G

 0  g ( y0 ) g  G
 a ( x , y 0 )  ( x ) x  E 42
Uniqueness :
If a ( x, y0 )  a ( x, y1 )  ( x ) x  E
then a ( x, y0  y1 )  0 x  E
Since E is reflexive,  x0  E with x0  1 s.t.
y1  y0  y1  y0 , x0
y1  y0
Suppose y1  y0 , then 1  , x0
y1  y0
y1  y0
then  J ( x0 )
y1  y0
y1  y0
then a ( x, )    0 
y1  y0 43
Variational Inequality
(Stampachia-Hartmam) p.1
E: reflexive Banach space
C: closed bounded convex set in E

f : C  E satisfies
(i) f is monotone i.e.
f ( x )  f ( y ), x  y  0  x, y  C
(ii) f is weakly continuous on each line
segment in C. 44
Variational Inequality
(Stampachia-Hartmam) p.2

Then there is y0  C such that

f ( y0 ), y0  x  0  x  C

45
Let h ( y , x )  f ( x ), x  y   x, y  C
Let S  C ,  : S  C ( S ) by  ( x )  h (, x )
Let q : C ( S )  R by q( h )  sup h ( s )
sS

s1 ,, s
For any p  Pf ( S )
1 ,,  
   
q( ( p ))  q   i ( si )   sup   i h ( y , si )
 i 1  yS i 1

 sup   i  f ( si ), si  y  
yS i 1

 sup   i  f ( y ), si  y   , sin ce f is monotone
yS i 1

 sup  f ( y ),   i si  y 
yS i 1

0 , by taking y ,   i si  S 46
i 1
q( ( p ))  0  p  Pf ( S )
by Mazru Olicz Thm
   C ( S ) with   q s.t.
 ( ( x ))  0  x  S
  ( h (, x ))  0 xS
By Lemma  y0 s.t.  ( h )  h ( y0 ) h  F (*)
Then for each x  S
 ( h (, x ))  h ( y0 , x )
  ( f ( x ), x    )  f ( x ), x  y0 
  ( f ( x ), x    f ( x ),  )  f ( x ), x    f ( x ), y0 
  f ( x ), x   ( f ( x ),  )  f ( x ), x    f ( x ), y0 
  ( f ( x ),  )  f ( x ), y0 
But from (*),  ( f ( x ),  )  f ( x ), y0 
Hence  ( f ( x ),  )  f ( x ), y0 
Therefore 0   ( h (, x ))  h ( y0 , x ) x  S
 f ( x ), x  y0  0 x  S 47
 f ( x ), y0  x  0 x  S
Now to show that  f ( y0 ), y0  x  0 x  C
For 0  t  1 , x  C
let zt  tx  (1  t ) y0
 f ( zt ), y0  zt  0
  f ( zt ), t ( y0  x )  0
  f ( zt ), y0  x )  0
By taking t  0, we have
 f ( y0 ), y0  x )  0

48
Applications of Mazm-
Orlich Theorem
Inequality after mixing of
functions

49
Theorem
Let S be an arbitary set.

f1,, f n : S  R

g1,, gm : S  R

The following two statements are equivalent:

50
( I )   ,   
n 1
 m 1

n m
i fi ( s )   j g j ( s)  s S
i 1 j 1

 ( II ) For any p  Pf ( S ), we have


min f i ( p )  max g j ( p )
1i  n 1 j  m

51
Appling Mazm  Orlicz Theorem
Let E1  R n , E2  R m
p1 ( x )  max xi  x   x1 ,, xn 
1i  n
p2 ( x )  max y j  y   y1 ,, ym 
1 j  n

 1 ( s )   f1 ( s ),, f n ( s )  sS
 2 ( s )   g1 ( s ),, g m ( s )  sS
( I )  ( II )
For any p  Pf ( S )
n
min f i ( p ) 
1i  n
 i f i ( p )
i 1
m
  j g j ( p ) , by ( I )
k 1
 max g j ( p ) 52
1 j  m
( II )  ( I )
For any p  Pf ( S ),
 p1 (  1 ( p ))   p1 (  f1 ( p ),, f n ( p ))
  max  f i ( p )
1i  n
 min f i ( p )
1i  n
 max g j ( p )
1 j  m

 p2 ( g1 ( p ),, g m ( p ))
 p2 ( 2 ( p ))
i.e.  p1 (  1 ( p ))  p2 ( 2 ( p )) 53
 the statement ( I ) of Mazm  Orlicz Theorem holds
   i  Ei * with  i  pi i  1,2 s.t
1 ( 1 ( s ))   2 ( 2 ( s )) s  S
i.e.  (1 ,,  n )  R n and (1 ,,n )  R m s.t
 n
  i xi  max xi for x   x1 ,, xn 
 i 1 1i  n
() m
  j y j  max y j for y   y1 ,, ym 

 j 1 1 j  m

n m
and  i f i ( s )   j g j ( s ) s  S
i 1 j 1

54
It remains to show that   n 1 and   m 1
n
By (  ) ,  i xi  max x1 ,, xn 
i 1
First take xi   k , k  0
x j 0 , if ji
  i k  0
 i  0  i  1,, n
Then take x  (1,,1), then
n
 i  1
i 1
Finally take x  ( 1,,1), then
n
  i  1
i 1
n
  i  1
i 1
n
Hence  i  1 i.e   n 1
i 1 55
m 1
Similarly,   
Appling Mazm  Orlicz Theorem
Let E1  R n , E2  R m
p1 ( x )  max xi  x   x1 ,, xn 
1i  n
p2 ( x )  max y j  y   y1 ,, ym 
1 j  n

 1 ( s )   f1 ( s ),, f n ( s )   s  S
 2 ( s )   g1 ( s ),, g m ( s )   s  S
Claim : If   1 ,,  n   E * , then   p1  1 ,,  n   n 1
[" " It is clear.
n
[" "   p1   i xi  max xi
i 1 1i  n
n
[ First take xi  1  i , we have  i  1
i 1
n
Then take xi  1  i , we have  i  1
i 1
n
Hence  i  1 .
i 1
Finally, take xi   k  0 and xj  0 if j i 56
then  i k  0  i  0
n n
 ,    n 1
m 1
s.t.  i f i ( s )  i gi ( s ) s  S
i 1 i 1
  linear functional
1  1 ,,  n  and  2  1 ,,m  s.t.
1  p1 and  2  p2
1 ( 1 ( s ))   2 ( 2 ( s ))  sS
 for any p  Pf ( S )
 ( p1 (  1 ( p )))  ( p2 ( 2 ( p )))
 for any p  Pf ( S )
 max  f i ( p )  max g j ( p )
1i  n 1 j  m

 for any p  Pf ( S )
min f i ( p )  max g j ( p ) 57
1i  n 1 j  m
Minimax Theorem of Von Neumann

For A  aij nm ,   , and   


n 1 m1

min max  aiji j  max min  aiji j


n 1 m1 ( i , j ) m1  n 1 ( i , j )

58
()
max  aiji j   aiji j    n 1 ,  m 1
m 1 ( i , j ) (i , j )
n 1 m 1
 max
 m 1
 aiji j  min

 aiji j n 1
   ,  
(i , j ) (i , j )
m 1
 min max
  
n 1 m 1
 aiji j  min

 aiji j n 1
  
(i , j ) (i , j )

 min max
  
n 1 m 1
 aiji j  max min  aiji j
   m 1 n 1
(i , j ) (i , j )

59
()
Apply Pr evious Theorem,
Let S  n 1  m 1 ,
 m

 f i ( , )   aij j
 j 1
 n
for ( , )  S
 gi ( , )   aiji

 i 1

( (1) , (1) ),, ( (  ) , (  ) )


Let p   Pf ( S )
1 ,, l
 
then p  (  k  (k )
,  k ( k ) )  S
k 1 k 1
 
Let ˆ   k  ( k )  n 1 and ˆ   k ( k )  m 1
k 1 k 1
60
then p  (ˆ ,ˆ )
m n
 fi ( p)   aijˆ j and g ( p )   aij ˆi
j 1 i 1
n
 min f i ( p ) 
1i  n
 ˆi f i ( p )
i 1
n m
  ˆi  aijˆ j
i 1 j 1
m n
 ˆ j  aijˆi
j 1 i 1
n
 max  aijˆi
1 j  m i 1
 max g j ( p )
1 j  m

 min f i ( p )  max g j ( p )
1i  n 1 j  m

i.e. statement ( II ) of previous Theorem


61
 statement ( I ) of previous Theorem
 
  0 , 0  n 1  m 1 s.t.
n m
 i f i ( s )
0

  j g j ( s)
0
 s S
i 1 j 1
n m n m
 i0
 aij j   j aiji
0
    ,    n 1 m 1

i 1 j 1 j 1 i 1

  ij i  j 
a  0
 ij i j
a   0
  ,   n 1  m 1
(i , j ) (i , j )

 max 
aiji  j
0

 minn 1 aiji j
0
  n 1
  ,     m 1
 ( i , j )
m 1
  ( i , j )

 min max
  n 1 m 1
 aiji j  max min  aiji j
   m 1 n 1
(i , j ) (i , j )
62
Lemma VI.1 (Riesz-Lemma)
Let

For any   fixed , apply Green’s second

1( x ,  ) in the domain


u  C ,  : C  domain, bounded
identity to u2and v 
 \ B   and then let   0 we have
 u ( x,  ) 
u    ( x,  )udx   ( x,  ) u  ds
 n x n x 63
64

You might also like