Вы находитесь на странице: 1из 14

Chapter 8

Bounded Linear Operator on a


Hilbert Space

8.1 Direct Sum and Projection


Let M and N be two linear subspaces of a linear space X. Then the sum

M + N := {x ∈ X | x = y + z with y ∈ M, z ∈ N } .

The direct sum

M ⊕ N := {x ∈ X | x can be written uniquely as x = y + z with y ∈ M, z ∈ N } .

M + N = M ⊕ N ⇐⇒ M ∩ N = {0}
Note: Let X = M ⊕ N , then

x = y + z = PM x + PN x = Ix, where I = PM + PN .

Also, for P = PM , ran P = M , ker P = N . Moreover, P 2 x = P (P x) = P x, i.e., P 2 = P .


Definition. Let X be a linear space. A linear map P : X → X is a projection if

P 2 = P.

Now, let’s see the relationship between direct sum and projection.
Proposition. If X = M ⊕ N , then there exists a projection P : X → X with

ran P = M, ker P = N.

(Why?) Take P = PM .
Proposition. If P : X → X is a projection on a linear space X. Then

X = ran P ⊕ ker P.

Proof. For all x ∈ X, x = P x + (x − P x). Note that x − P x ∈ ker P since P (x − P x) = P x − P 2 x = 0. So


x = ran P ⊕ ker P .
If x ∈ ran P + ker P , then x = P y, P x = 0 and thus

0 = P x = P 2 y = P y = x.

So ran P ∩ ker P = {0}, X = ran P ⊕ ker P .

1
8.2 Orthogonal Projection
Definition. Let M be a closed subspace of a Hilbert space H. Then

H = M ⊕ M⊥ (orthogonal direct sum).

In this case, P : H → M is the orthogonal projection.


Definition. A linear map P : H → H on a Hilbert space H is an orthogonal projection if

P2 = P and P ∗ = P.

( P is self-adjoint, i.e., P ∗ = P , hP x, yi = hx, P yi for all x, y ∈ H. )


Proposition. If M is a closed subspace of a Hilbert space H, then

P = PM : H → H

is an orthogonal projection with


ran P = M, ker P = M ⊥ .
Proof. We only need to prove that P is self adjoint. For all x, y ∈ H,

x = xM + xM ⊥ , y = yM + yM ⊥ ,

hP x, yi = hxM , yi = hxM , yM i = hx, yM i = hx, P yi.


Thus, P is self-adjoint.
Proposition. If P is an orthogonal projection on a Hilbert space H, then ran P is closed, and

H = ran P ⊕ ker P

is the orthogonal direct sum of ran P and ker P .


Proof. P is a projection on H implies
H = ran P ⊕ ker P
as direct sum. But if x = P y ∈ ran P , z ∈ ker P , then, since P is self-adjoint,

hx, zi = hP y, zi = hy, P zi = 0.

So ran P ⊥ ker P and thus ran P is closed (because it is a perpendicular complement of some set.)
Proposition. Let P : H → H be an orthogonal projection. Then

(1) If P is nonzero, then kP k = 1.


(2) I − P : H → H is also an orthogonal projection.
Proof. (Why?) (1) For all x ∈ H,
2
kP xk = hP x, P xi = hx, P 2 xi = hx, P xi ≤ kxk kP xk .
° °
So kP xk ≤ kxk and thus kP k ≤ 1. If P 6= 0, then there exists x ∈ H with y = P x 6= 0 and kP yk = °P 2 x° =
kP xk = kyk. So kP k = 1.
(2) Observe that
(I − P )2 = I − 2P + P 2 = I − 2P + P = I − P.
Also, for all x, y ∈ H,

h(I − P )x, yi = hx − P x, yi = hx, yi − hP x, yi = hx, yi − hx, P yi = hx, (I − P )yi.

2
Example. Let H = L2 (R), M, N be the collections of all odd and even functions in L2 (R) respectively.
Then
L2 (R) = M ⊕ N.
(Why?) For all f ∈ L2 (R), f = feven + fodd where

f (x) + f (−x) f (x) − f (−x)


feven = , fodd = .
2 2
So H = M + N , M ∩ N = {0} since the integral of a odd function, obtained by an even function times an
odd function, is zero.
Example. Let (X, Σ, µ) measure space, A ∈ Σ be measurable set. Let

P = L2 (X, µ) → L2 (X, µ), f 7→ f χA .

Then
P 2 f = f χA χA = f χA = P f
Z
hP f, gi = f χA g dµ = hf, P gi.

So it is an orthogonal projection.

8.3 The Dual of a Hilbert Space


Recall: Let X be a linear space. Then, the dual space

X ∗ = {bounded linear functions on X} = {φ : X → C | φ bounded, linear } .

φ bounded means
kφ(x)k
kφk = sup < ∞.
x6=0 kxk

X = B(X, C): Banach space.
If X is Banach, then X ⊆ X ∗∗ . Now, if X is Hilbert, we’ll show that X = X ∗ .
Let H be a Hilbert space. For any y ∈ H, define

φy : H → C, φy (x) = hy, xi for all x ∈ H.

1. φy is linear.

2. |φy (x)| = |hy, xi| ≤ kyk kxk, that is, kφy k ≤ kyk. On the other hand, |φy (y)| = |hy, yi| = kyk kyk. To
conclude,
kφy k = kyk , φy ∈ H∗ .

So, we get a map


J : H → H∗ , y 7→ φy .
2
J is 1-1 because φy = 0 implies hy, xi = 0 for all x ∈ H, i.e., hy, yi = kyk = 0, y = 0. Now, we’ll show that
J is onto. That is, for any φ ∈ H∗ , there exists y ∈ H such that φ = φy .
Example. Let H = R3 , φ be linear and bounded. Then φ can be written as

φ(x, y, z) = ax + by + cz = (a, b, c) · (x, y, z).

Theorem (Riesz representation theorem). If φ is a bounded linear functional on a Hilbert space H (i.e.,
φ ∈ H∗ ), then there is a unique vector y ∈ H such that

φ(x) = hy, xi

for all x ∈ H. Furthermore, kφk = kyk.

3
Remark. If y exists, for all x ∈ ker(φ),
0 = φ(x) = hy, xi.
So y ∈ (ker(φ))⊥ .
Proof. If φ = 0, then y = 0. So we may assume that φ 6= 0. In this case, ker(φ) is a proper closed subspace
(the kernel is closed because φ is linear and bounded) of H. So

H = ker(φ) ⊕ ker(φ)⊥

as direct sum. Pick any z ∈ (ker(φ))⊥ with z 6= 0, then for all x ∈ H,


µ ¶
φ(x) φ(x)
x= z+ x− z .
φ(z) φ(z)
φ(x)
Note that n = x − φ(z) z ∈ ker(φ) because

φ(x)
φ(n) = φ(x) − φ(z) = 0.
φ(z)

So
φ(x) φ(x) 2 φ(x) 2
hz, xi = hz, z + ni = kzk + hz, ni = kzk ,
φ(z) φ(z) φ(z)
and thus
φ(z) φ(z)
φ(x) = 2 hz, xi = hy, zi, where y = 2 z.
kzk kzk
So φ = φy .
For the uniqueness, Suppose φ = φy + φye for some y, ye ∈ H. Then for all x,

hy − ye, xi = hy, xi − he
y , xi = φ(x) − φ(x) = 0.
2
so hy − ye, y − yei = ky − yek = 0 and thus y = ye.
φ(z) kzk2
Remark. Assume φ 6= 0. For all z ∈ (ker(φ))⊥ , y = kzk2
z. So z = y, i.e.,
φ(z)

(ker(φ))⊥ = Span {y} .

Remark. Geometric interpretation: For all φ ∈ H∗ , we consider level sets of

φ = φy = {φ(x) = c} = {hy, zi = c} ,

parallel hyperplanes.
Now, J is 1-1, onto, and isometry (kφy k = kJyk = kyk). So we may identify H with H∗ , and H is
self-dual.

8.4 The Adjoint of an Operator


Proposition. Let A : H → H be a bounded linear operator on a Hilbert space H. Then there exists a
unique bounded linear operator A∗ : H → H such that

hx, Ayi = hA∗ x, yi

for all x, y ∈ H.
Proof. For each fixed x ∈ H, the map
φx : H → R
defined
φx (y) = hx, Ayi

4
is linear and bounded. Indeed, since A is bounded, by the Cauchy-Schwarz inequality,

kφx (y)k = |hx, Ayi| ≤ kxk kAyk ≤ kxk kAk kyk .

So kφx k ≤ kxk kAk. By the Riesz representation theorem, there exists a unique z ∈ H such that

φx (y) = hx, yi

for all y ∈ H. Se set A∗ x = z, then


hx, Ayi = hA∗ x, yi.
A∗ is linear because

hA∗ (ax1 + bx2 ), yi = hax1 + bx2 , Ayi


= ahx, Ayi + bhx, Ayi
= ahA∗ x, yi + bhA∗ x, yi
= haA∗ x1 + bA∗ x2 , yi

for all y ∈ H. By the uniqueness in the Riesz representation theorem,

A∗ (ax1 + bx2 ) = aA∗ x1 + bA∗ x2 .

A∗ is bounded because
2
kA∗ xk = hA∗ x, A∗ xi = hx, AA∗ xi ≤ kxk kAA∗ xk ≤ kxk kAk kA∗ xk .

So kA∗ xk ≤ kxk kAk, that is, kA∗ k ≤ kAk and thus A∗ is bounded.
Proposition. Some simple properties of adjoint operators.
1. A∗∗ = A (i.e., (A∗ )∗ = A).
(Why?) Note that
hA∗∗ x, yi = hx, A∗ yi = hA∗ y, xi = hy, Axi = hAx, yi.
So A∗∗ x = Ax for all x, and thus A∗∗ = A.
2. kA∗ k = kAk.
(Why?) Observe that
kA∗ k ≤ kAk = k(A∗ )∗ k ≤ kA∗ k .

3. (AB)∗ = B ∗ A∗ .
(Why?)
h(AB)∗ x, yi = hx, AByi = hA∗ x, Byi = hB ∗ A∗ x, yi.

Example. H = Rn . Linear map: Ax, A = [aij ]n×n , hx, yi = xT y, hx, Ayi = xT Ay = (AT x)T y = hAT x, yi,
adjoint map: AT x.
T
H = Cn . Linear map: Ax, A = [aij ]n×n , hx, yi = xT y, adjoint map: A∗ x where A∗ = A , the Hermitian
conjugate matrix.
Example. H = l2 (N), Right shift operator

S : (x1 , x2 , . . .) → (0, x1 , x2 , . . .).

Left shift operator


T : (x1 , x2 , . . .) → (x2 , x2 , . . .).
kSk = kT k = 1,

X
hx, Syi = h(x1 , x2 , . . .), (0, y1 , y2 , . . .)i = xi+1 yi
i=1
= h(x2 , x3 , . . .), (y1 , y2 , . . .)i = hT x, yi.

So S ∗ = T , T ∗ = S.

5
Example. H = L2 (Rn ). Integral operator
Z
Kf (x) = k(x, y)f (y)dy, for all f ∈ L2 (Rn ),
Rn

where k(x, y) ∈ L2 (Rn × Rn , C).


Z µZ ¶
hKf (x), g(x)i = k(x, y)f (y) dy g(x) dx
n Rn
ZR Z
= k(x, y)f (y)g(x) dxdy
n Rn
ZR Z
= f (y) k(x, y)g(x) dxdy
Rn Rn
Z Z
= f (x) k(y, x)g(y) dydx
Rn Rn
Z
∗ ∗
= hf, K yi, where K y= k(y, x)g(y) dy.
Rn

T
cf. A∗ = A .
Theorem. If A : H → H is a bounded linear operator, then

H = ran A ⊕ ker A∗

as orthogonal direct sums. That is,

ran A = (ker A∗ )⊥ , ker A = (ran A∗ )⊥ .

Proof. For all x ∈ ran A, there exists y ∈ H such that x = Ay. For any z ∈ ker A∗ ,

hx, zi = hAy, zi = hy, A∗ xi = hy, 0i = 0.

So ran A ⊥ ker A∗ , ran A ⊆ (ker A∗ )⊥ .


On the other hand, for all x ∈ (ran A)⊥ , for all y ∈ H,

hA∗ x, yi = hx, Ayi = 0.

So A∗ x = 0 and x ∈ ker A∗ . So (ran A)⊥ ⊆ ker A∗ , and thus

ran A = (ran A)⊥⊥ ⊇ (ker A∗ )⊥ .

So ran A = (ker A∗ )⊥ . Apply to A∗ ,



ran A∗ = (ker A)⊥ , ker A = (ran A∗ ) = (ran A∗ )⊥ .

Application: (Fredholm alternative) Solve a linear equation

Ax = y (*)

where A : H → H is a bounded linear operator. Solution of (*) exists ⇐⇒ y ∈ ran A.


(a) If ran A is closed, then solution of (*) exists ⇐⇒ y ∈ (ker A∗ )⊥ .
(b) If ran A is closed and ker A∗ = {0}, then solution of Ax = y exists for all y ∈ H.
Note: ker A∗ = {0} ⇐⇒ the solution of the adjoint problem A∗ x = y is unique.
ran A is closed ⇐ if we have an estimate of the form c kxk ≤ kAxk.

6
8.5 Self-Adjoint and Unitary Operator
8.5.1 Self-Adjoint Operator
Definition. A bounded linear operator A : H → H on a Hilbert space H is self-adjoint is A = A∗ .
That is,
hx, Ayi = hAx, yi for all x, y ∈ H.
Example. H operator adjoint self-adjoint
Rn Ax AT x A = AT symmetric
Cn Ax A∗ x A = A∗ Hermitian
R R
L (Rn )
2
Kx = Rn k(x, y)f (y)dy Rn
k(y, x)f (y)dy k(x, y) = k(y, x)
Example. Let P : H → H be an orthogonal projection. That is,

P 2 = P, hP x, yi = hx, P yi.

So P is self-adjoint.
Example. Let A be a bounded linear operator on H. Then A∗ A is self-adjoint because

(A∗ A)∗ = A∗ (A∗ )∗ = A∗ A.

Note that AA∗ is also self-adjoint.


Lemma. If A is a bounded self-adjoint operator on a Hilbert space H. Then

kAk = sup |hx, Axi| .


kxk=1

Proof. (≥) For all x ∈ H with kxk = 1,

|hx, Axi| ≤ kxk kAxk ≤ kxk kAk kxk = kAk .

So
kAk ≥ sup |hx, Axi| .
kxk=1

(≤) Let α = supkxk=1 |hx, Axi|. Want to show kAk < α, that is, kAxk ≤ α whenever kxk ≤ 1. If Ax = 0,
Ax
then kAxk = 0 ≤ α. If Ax 6= 0, then let y = kAxk , thus kyk = 1,

2
Ax kAxk
hy, Axi = h , Axi = = kAxk ∈ R.
kAxk kAxk

So
hy, Axi = hy, Axi = hAx, yi = hx, Ayi,

hx + y, A(x + y)i − hx − y, A(x − y)i = hx, Axi + hx, Ayi + hy, Axi + hy, Ayi
− hx, Axi + hx, Ayi + hy, Axi − hy, Ayi
= 4hy, Axi = 4 kAxk .

Thus,

4 kAxk ≤ |hx + y, A(x + y)i| + |hx + y, A(x + y)i|


2 2
≤ α kx + yk + α kx − yk
2 2
= α(2 kxk + 2 kyk ) = 4α,

So kAxk ≤ α, i.e., kAk ≤ α.


To conclude kAk = α = supkxk=1 |hx, Axi|.
2
Corollary. If A is a bounded linear operator on H, then kA∗ Ak = kAk .

7
Proof.

kA∗ Ak = sup |hx, A∗ Axi| = sup |hAx, Axi|


kxk=1 kxk=1
à !2
2 2
= sup kAxk = sup kAxk = kAk
kxk=1 kxk=1

° °
Corollary. If A is self-adjoint, then °A2 ° = kAk .
2

Corollary. If P is a nonzero orthogonal projection, then kP k = 1.


° °
Proof. Observe that kP k = °P 2 ° = kP k . So kP k = 0 or 1, But P is nonzero. Thus, kP k = 1.
2

8.5.2 Unitary Operator


Definition. A linear map U : H1 → H2 between two (complex) Hilbert spaces is unitary if

(1) U is bijective (1-1 and onto),


(2) hU x, U yiH2 = hx, yiH1 for all x, y ∈ H1 .
Two Hilbert space are isomorphic if there is a unitary map between them.
Remark. Condition (2) is equivalent to

kU xkH2 = kxkH1 for all x ∈ H.

This follows from the polarization formula


µ µ° °2 ° x °2 ¶¶
1 2 2 °x ° ° °
hx, yi = kx + yk − kx − yk + i ° + y ° − ° − y ° .
4 i i

Remark. A linear map U : H → H is unitary if and only if U ∗ U = U U ∗ = I.


(Why?)
(⇐) U ∗ U = U U ∗ = I, then U is 1-1 and onto and hU x, U yi = hx, U ∗ U yi = hx, yi.
(⇒) Since hU x, U yi = hx, yi, we have hx, U ∗ U yi = hx, yi for all x.
That is, U ∗ U = y for all y and thus U ∗ U = I.
Example. Let H1 , H2 be two Hilbert space of the same dimension. Let {uα }α∈A , {vα }α∈A be an orthonor-
mal basis of H1 and H2 respectively. Then
X X
U : H1 → H2 , xα uα 7→ x α vα
α∈A α∈A

is unitary. So H1 is isomorphic to H2 .
For example, X n o
L2 (T) → l2 (Z), fbn einx 7→ fbn .
n∈Z
n∈Z

More generally, let λα = eiθα , θα ∈ R. Then |λα | = 1,


X X
U : H1 → H2 , xα uα 7→ xα λα vα
α∈A α∈A

is also unitary. Thus, U is 1-1, onto, and


X
hx, yi = xα yα
α∈A
X X
hU x, U yi = xα λα yα λα = xα yα
α∈A α∈A

8
8.6 Weak Convergence on a Hilbert Space
Recall xm → x is a Banach space X if φ(xn ) → φ(x) as n → ∞. for every φ ∈ X ∗ .
Now, assume H is a Hilbert space. By Riesz representation theorem, H∗ = H, for all φ ∈ H∗ , there exists
y ∈ H such that φ = φy where φy (x) = hy, xi.
Definition. A sequence {xn } is a Hilbert space H converge weakly to x ∈ H if

hy, xn i → hy, xi

for any y ∈ H. Usually we write xn * x.


Note that if xn → x, then xn * x. (Why?) For all y ∈ H,

|hxn , yi − hx, yi| = |hxn − x, yi| ≤ kxn − xk kyk → 0.

So hxn , yi → hx, yi and thus xn * y.


But xn * x ; xn → x.

Example. Let {en }n=1 be an orthonormal set in a Hilbert space H. Then we can have en * 0 but en 9 0.

Proof. For all y ∈ H, by Bessel’s inequality,


X 2 2
|hen , yi| ≤ kyk < ∞.
n

So hen , yi → 0 as n → ∞. and thus en * 0. However, ken − em k = 2 if n 6= m. So en 9 0 (or ken k = 1,
lim ken k = 1 6= 0).

Theorem (Uniform boundness principle / Banach-Steinhaus). Assume {φn } is a set of linear bounded
functionals on a Banach space X. If

sup |φn (x)| < ∞ for any x ∈ X, (*)


n

then
sup kφn k < ∞. (**)
n

Remark. Note that (*) means {φn (x)} is bounded for every x. So there exists Mx > 0 such that

|φn (x)| ≤ Mx kxk .

(**) means {kφn k} is bounded. So there exists M > 0 such that

kφn (x)k ≤ M kxk .

That is, pointwise bounded implies uniform bounded. Also, if supn kφn k = ∞. then there exists x0 such
that supn kφn (x)k = ∞.
Proposition. If xn * x in a Hilbert space H, then {xn } is bounded (i.e., there exists M such that kxn k ≤ M
for all n).
Proof. For all n, let φn = φxn , i.e., φn (y) = hxn , yi, kφn k = kxn k. For any y ∈ H, since {φn (y)} is convergent
to hx, yi it is a bounded sequence. By the uniform boundness principle, {kφn k} is bounded. So {kxn k} is
bounded.
Q: Given a sequence {xn } in H, how to show xn * x? Method 1: Show that hxn , yi → hx, yi FOR ALL
y ∈ H.
Theorem. Assume that {xn } is a sequence in a Hilbert space H, and D is a dense subset of H. Then
xn * x if and only if
1. kxn k ≤ M for some constant M .
2. hxn , yi → hx, yi for all y ∈ D.

9
Proof. (⇒) Done.
(⇐) If z ∈ H, then for any ² > 0, there exists y ∈ D such that kz − yk < ². Since hxn − x, yi → 0, there
exists N such that
|hxn − x, yi| < ² whenever n ≥ N.
So

|hxn − x, zi| ≤ |hxn − x, z − yi| + |hxn − x, yi|


≤ kxn − xk kz − yk + ²
≤ (M + kxk)² + ² = (M + kxk + 1)²

So hxn , zi → hx, zi for all z ∈ H.


If {eα } is an orthonormal basis of H, then D = Span {eα } is dense in H. Thus Method 2 yields that
hxn , yi → hx, yi for all y ∈ D if and only if hxn , eα i → hx, eα i for all α.
To conclude, xn * x if and only if
(1) {xn } is bounded,
(2) hxn , eα i → hx, eα i for any α, where {eα } is an orthonormal basis of H,
Note that condition (2) means that coordinates of xn converges to coordinates of x, and condition (2) alone
with out (1) does not imply weak convergence.
Example. H = l2 (N), en = (0, 0, . . . , 0, 1, 0, . . .), xn = nen = (0, 0, . . . , 0, n, 0, . . .). For each coordinate (m
fixed),
hxn , em i = nδnm → 0 as n → ∞
but xn is not weak convergent to 0.
Let y = (1, 21p , 31p , . . . , n1p , . . .),
X X 1 1
yn2 = <∞ if p > .
n n
n2p 2

So y ∈ l2 (N). But
1 1
hxn , yi = n · = n1−p 9 0 = h0, yi if < p < 1.
np 2
So xn not weak to 0.
Here are some typical ways that weak convergent sequence fails to converge strongly: oscillations, con-
centration, escape to infinity....
© √ ª
Example. H = L2 [0, 1]. en = 2 sin(nπx) is an orthonormal basis of H. Then en * 0 but en 9 0. So
for any f ∈ L2 [0, 1],
√ Z 1
hen , f i = 2 sin(nπx)f (x) dx → 0 as n → ∞.
0
Here, en oscillates more and more rapidly as n → ∞.
(√
n if 0 ≤ x ≤ n1
Example. fn = 1
. fn * 0 in L2 [0, 1] but fn 9 0 because kfn k = 1.
0 n < x ≤ 1
(Why fn * 0?) Pick D = C ∞ [0, 1] dense in L2 [0, 1]. For all f ∈ D,
Z 1 Z 1 √ R n1
n √ n 0 f (x)dx
hfn , f i = fn f dx = nf (x)dx = 1 → 0 · f (0) = 0.
0 0 n n

So fn * f .
Example. H = L2 (R), fn = χ[n,n+1] . Thus kfn k = 1 but fn * 0. Pick D = Cc∞ (R) is dense in H. For all
f ∈ Cc∞ (R). For all f ∈ Cc∞ (R),
Z n+1
hfn , f i = f (x)dx → 0 as n → ∞.
n

So fn * 0 but fn 9 0.

10
Proposition.
(1) If xn * x, then
kxk ≤ lim inf kxn k .
n→∞

(2) xn → x if and only if


xn * x and kxk = lim kxn k .
n→∞

Proof. (1) If xn * x,
2
kxk = hx, xi = limhx, xn i ≤ lim inf kxk kxn k .
n n→∞
So kxk ≤ lim inf n→∞ kxn k.
(2) If xn → x, then kxn k → kxk because k·k is continuous. If kxn k → kxk and xn * x, then
2 2 2
kxn − xk = hxn − x, xn − xi = kxn k − hx, xn i − hxn , xi + kxk
2 2
→ kxk − hx, xi − hx, xi + kxk = 0
as n → ∞.
Proposition. Every bounded sequence in a Hilbert space has a weakly convergent subsequence. That is, if
kxn k ≤ M for all n, then there exists a subsequence {xnk } such that xnk * x for some x ∈ H.
Remark. Let {xn } be a bounded sequence in H.
1. If dim H < ∞, we have H = Cn . By the Bolzano-Weierstrass theorem, {xn } as a (strongly) convergent
subsequence.
2. If dim H = ∞, {xn } may fail to have a strongly convergent subsequence. e.g., xn = en an orthonormal
basis of H. then {xn } has no convergent subsequence.
3. This proposition says {xn } always have a weakly convergent subsequence even dim(H) = ∞.
4. A set is weakly precompact ⇐⇒ A is bounded.

Proof. (⇐) If A is bounded, for any sequence {xn } in A, {xn } is a bounded sequence. So it has a
weakly convergent subsequence. thus A is weakly precompact.
(⇒) If A is precompact and suppose A is not bounded, then there exists a sequence {xn } such that
kxn k ≥ max {kxi k + 1, n} .
1≤i≤n−1

Then {xn } has no bounded subsequence. Since A is weakly precompact, {xn } has a weakly convergent
subsequence {xnk }. Thus {xnk } bust be bounded, a contradiction.

Proof of Proposition. We’ll prove the result only for a separable Hilbert space. Let {en }n=1 be an orthonor-
mal basis of H. For each xk , X
xk = hen , xk ien .
n
Thus, X
2 2
kxk k = |hen , xk i| ≤ M 2 .
n
So |hen , xk i| ≤ M for all n, k. Now, we may use a diagonal argument to see that {xk } has a subsequence
{xk,k } such that hen , xk,k i is convergent as k → ∞ for all n = 1, 2, . . ..
Indeed, since he1 , xk i is bounded in C, so {xk } has a subsequence {x1,k } such that he1 , x1,k i converges as
k → ∞. Since he2 , x1,k i is bounded in C, so {x1,k } has a subsequence {x2,k } such that he1 , x2,k i, he2 , x2,k i
converges as k → ∞. Continuing in this way, we get a subsequence {xnk } such that hej , xn,k i converges as
k → ∞ for all j = 1, 2, . . .. Now, taking the diagonal subsequence {xk,k } of {xk }, we see hej , xk,k i converges
as k → ∞ for j = 1, 2, 3, . . ..
Let an = limk→∞ hen , xk,k i ∈ C. Since, with Fatou’s lemma,
X 2
X 2
X 2 2
|an | = lim |hen , xk,k i| ≤ lim inf |hen , xk,k i| = lim inf kxk,k k ≤ M 2 .
k k k
n n n
P
So x = n an en ∈ H; hen , xk,k i → an = hen , xi and kxk,k k is bounded, se have xk,k * x.

11
Corollary (Banach-Alaoglu). The closed unit ball of a Hilbert space is weakly compact.
(Why?) For any {xn } in B1 , kxn k ≤ 1. So {xn } has a weakly convergent subsequence {xnk } such that
xnk * x, and thus
kxk ≤ lim inf kxnk k ≤ 1, x ∈ B1 .
k→∞

So B1 is sequentially weakly compact which also implies B1 is weakly compact.

8.7 Application
Consider a minimization problem
min f (x) (*)
x∈K

where f : K → R. Question: For which f , K does (*) have a minimum?


Remark. If K is a compact metric space and f is lower-semicontinuous, then (*) has a minimizer.
Theorem. If K is a weakly closed, bounded subset of a Hilbert space, and f is weakly lower-semicontinuous
in the sense that
f (x) ≤ lim inf f (xn )
n→∞

for every sequence {xn } in K such that xn * x, then (*) has a minimizer.
Remark. In metric space,
compact = precompact & closed.
If X is complete,
precompact = totally bounded.
In Hilbert space,
weakly compact = weakly precompact & weakly closed.
cf:
totally bounded ⇒ bounded.
However,
closed ⇐ weakly closed.
convergent sequence ⊆ weakly convergent sequence.
(Why?) Assume K is weakly closed. For any xn → x with xn ∈ K, xn * x. But, K is weakly closed. So
x ∈ K and thus K is closed.
Similarly,
f weakly lower-semicontinuous ⇒ f lower-semicontinuous.
Proof. Step 1: Taking a minimizing sequence {xn } such that

f (xn ) → inf f (x)


x∈K

since the infimum of f exists.


Step 2: Since K is bounded, {xn } has a weakly convergent subsequence such that xnk * x∗ . for some x∗ .
because K is weakly closed.
Step 3: Since f is weakly lower-semicontinuous,

f (x∗ ) ≤ lim inf f (xn ) = inf f (x)


k→∞ x∈K

and x∗ ∈ K, So f (x∗ ) = min f (x).


x∈K

Theorem. Assume K is a (strongly) closed, convex, bounded subset of a Hilbert space, and f : K → R is
(strongly) lower-semicontinuous, convex function on K. Then (*) has a minimizer. If f is strictly convex,
then (*) has a unique solution.

12
Definition. A subset C of a linear space is convex if

tx + (1 − t)y ∈ C

for all x, y ∈ C, t ∈ [0, 1].


Definition. A function f : C → R is a convex function on a convex set C if
¡ ¢
f tx + (1 − t)y ≤ tf (x) + (1 − t)f (y)

for all x, y ∈ C and t ∈ [0, 1]. f is a strictly convex function if we get “<” when x 6= y, t ∈ (0, 1).
1 2
Example. 2 kxk is strictly convex. For all φ ∈ H∗ , φ is convex.
Remark. Let {x1 , x2 , . . . , xn } be n vectors in a linear space X. We may use it to generater a convex set
\
C= {convex set E of X such that E ⊇ {x1 , x2 , . . . , xn }}
( n n
)
X X
= y= λk xk , λ = 1, λk ≥ 0.
k=1 k=1

Here we say y is a convex combination of {x1 , x2 , . .P


. , xn }.
Recall that xn → x ⇒ xn * x; xn * x ⇒ yn = λk xk → x.
Theorem (Mazur). If xn * x in a Hilbert space, then there is a sequence {yn } of finite convex combination
of {xn } such that yn → x.
Proof. By replacing xn by xn − x, we may assume xn * 0. Thus,

hxn , yi → h0, yi = 0 for all y.

Also, {xn } is bounded by a constant M > 0. Pick xn1 = x1 . Since hxn , xn1 i → 0 as n → ∞, there exists
n2 > n1 such that
|hxn2 xn1 , i | ≤ 1.
Given n1 , n2 , . . . , nk , since
hxn , xni i → 0 as n → ∞ for i = 1, 2, . . . , k,
there exists nk+1 > nk such that
¯ ¯ 1
¯hxn , xni i¯ ≤ .
k+1
k
1
Pk
Let yk = k i=1 xni , Then
k k
2 1 XX
kyk k = hxn , xnj i
k 2 i=1 j=1 i
k k i−1
1 X 2 X X ¯¯ ¯
hxni , xnj i¯
2
≤ 2
kxni k + 2
k i=1 k i=1 j=1
k i−1
km2 2 XX 1
≤ +
k2 k 2 i=1 j=1 i
km2 2k m2 + 2
≤ = →0 as k → ∞.
k2 k2 k

Proof of the previous theorem.


(1) Strongly closed, convex ⇒ weakly closed.
(2) Strongly lower-semicontinuous, convex ⇒ weakly lower-semicontinuous (HW)
As for (1), if xn * x, xn ∈ K, by Mazur theorem, we have yn → x. (Note that yn ∈ K because K is
convex.) So x ∈ K because K is strictly closed. So K is weakly closed.

13
(3) Strongly convex ⇒ unique solution. Assume f achieve it minimum at x∗ , y ∗ with x∗ 6= y ∗ . Then

x∗ + y ∗ 1 1
f( ) < f (x∗ ) + f (y ∗ ) = f (x∗ ).
2 2 2
This contradicts with x∗ being the minimum.

Theorem. Assume
(1) C is strongly closed, convex subset of H (here C may be unbounded),

(2) f : C → R is coercive strongly lower-semicontinuous convex function.


Then (*) has a minimum. Here coercive means

lim f (x) = ∞.
kxk→∞

14

Вам также может понравиться