mal basis.
Proof.
Call the space X. Since it is not 0, it contains a nonzero vectorx.
The set consisting solely of
x/llxll
is orthonormal. Now order the family of all orthonormal subsets of X in the natural way (by inclusion). In order to use Zorn's Lemma, one must verify that each chain of orthonormal sets has an upper bound. Let C be such a chain, and putA'
=U{A : A
E C} . It is obvious thatA'
is an upper bound for C, but isA'
orthonormal? Takex
and y inA'
suchthat
x ;6 y.
Sayx
EAl
E C andy
EA2
E C. Since C is a chain, eitherAl
CA2
or
A2
CAI.
Suppose the latter. Thenx, y
EAI.
SinceAl
is orthonormal,(x, y) = O.
Obviously,IIxll =
1. HenceA'
is orthonormal. • Theorem 6. The Orthonormal Basis Theorem.For an or
thonormal family Iu;! (not necessarily /inite or countable) in a Hilbert space
X,the following properties are equivalent:
a.
Iu;! is an orthonormal basis for
X.b.
Ifx
E Xand x .l Ui for all i, then x =
o.c.
For each x
E X,x = L(X, Ui)Ui.
d.
For each x and y in
X,(x, y)
=L(X, Ui)(y, Ui)'
e .
For each x in
X,IIxll2 = L I (x, Ui)j2 . (Parseval Identity)
Proof.
To prove that a implies b, suppose that b is false. Letx ;6 0
andx
.1Ui
for alli.
Adjoinx/ llxll
to the familyIu;)
to get a larger orthonormal family. Thus the original family is not maximal and is not a basis.To prove that b implies c, assume b and let
x
be any point in X. Let Y= L (X, Ui)Ui'
By Bessel's inequality (Theorem4),
we haveBy Theorem
2,
the series definingy
converges. (Here the completeness of X is needed.) Then straightforward calculation (as in the proof of Theorem3)
shows thatx - y
.1Ui
for alli.
By b,x - y = O.
74 Chapter 2 Hilbert Spaces
To prove that c implies d, assume c and writ.eStraightforward calculation then yields To prove that d implies e , assume d
(x,y)
and let= L(X,Ui)(y,Ui)' y
=x
i n d . The result is the assertion in e.To prove that e implies a, suppose that a is false. Then
lUi]
is not a maximal orthonormal set. Adjoin a new element,x,
to obtain a larger orthonormal set.Then
1 = IIxl12
1=L I(x, ui)12
=0,
showing that e is false. • Example 1. One orthonormal basis ine2
is obtained by definingunU) = 6nj.
Thus
Ul
=[1, 0, 0,
. ..
J ,U2 = [0, 1, 0,
. . . J , etc.To see that this is actually an orthonormal base, use the preceding theorem, in particular the equivalence of a and
b.
Supposex
Ee2
and(x, un)
=0
for all n.Then
x (
n)
=0
for all n, andx =
o. •Example 2. An orthonormal basis for
L2[0, 1]
is provided by the functionsun(t) = e2rrint,
where n E Z. One verifies the orthonormality by computing the appropriate integrals. To show that[un]
is a base, we use Partb
of Theorem6.
Let
x
EL2[0, 1]
andx
1=O.
It is to be shown that(x,un)
1=0
for some n. Sincethe set of continuous functions is dense in
L2,
there is a continuousy
such thatIlx - YII < Ilxli/5.
ThenlIyll
�Ilxll - llx - YII > �llxll·
By the Weierstrass Approximation Theorem, the linear span of[unJ
is dense in the spaceC[O, lJ,
furnished with the supremum norm. Select a linear combination
p
of[un]
suchthat
�lIxli. lip - YII
Then 00< IIxli/5.
Thenlip - YII < IIxll/5.
HenceIIpl! > IIYII - IIY -pll >
l(x,p)1
�l(p,p)I - I(y -p,p)I - I(x - y,p)1
�
IIpll2 - IIY - pll IIpil - IIx - yll IIpll > 0
Thus it is not possible to have
(x, un) = 0
for all n. • Recall that we have defined the orthogonal projection of a Hilbert spaceX
onto a closed subspaceY
to be the mappingP
such that for eachx
EX, P x
is the point of
Y
closest tox.
Theorem 7. The Orthogonal Projection Theorem.
The ortllOgonal projection P of a Hilbert space X onto a closed subspace Y has tlJese properties:
a.
It is well-defined; i.e., Px exists and is unique in Y.
b. It is surjective, i.e., P(X)
=Y.
c.
It is linear.
d. If e.
x
--Y is not Px .1 Y for all 0 (the zero subspace), then x. IIpll = 1.
Section
2.2Orthogonality and Bases
f.
P is Hermitian; i.e., (Px, w) = (x, Pw) [or all x and w.
g.
If[Yd is an orthonormal basis [or Y, then Px = "L(X,Yi)Yi.
h.
P is idempotent; i.e., p2 = P.
i.
Py = Y [or all Y
EY. Thus PlY =
fl"j.
IIxll2
=IIpxll2
+llx - Px1J2.
Proof. This is left to the problems.
75
• The Gram-Schmidt process, familiar from the study of linear algebra, is an algorithm for producing orthonormal bases. It is a recursive process that can be applied to any linearly independent sequence in an inner-product space, and it yields an orthonormal sequence, as described in the next theorem.
Theorem 8. The Gram-Schmidt Construction.
Let
[VI, V2, V3,
... J be a linearly independent sequence in an inner product space. Having set UI = vI/llvlll, define recursively
Vn - "L (vn, Ui)Ui n-I i=1 Un
==---'-n--:"'I IIVn - "L (Vn, Ui)udl
i=1 ----
n =
2, 3, . . .
Then [UI' U2, U3, .
..J is an orthonormal sequence, and [or each
n, span{u I
'U2, .. .
,u n}
== span{VI, V2, · .. , vn}.
Notice that in the equation describing this algorithm there is a normalization process: the dividing of a vector by its norm to produce a new vector pointing in the same direction but having unit length. The other action being carried out is the subtraction from the vector
Vn
of its projection on the linear span of the orthonormal set presently available,UI, U2,'" ,Un-I.
This action is obeying the equation in Theorem3,
and it produces a vector that is orthogonal to the linear span just described. These remarks should make the formulas easy to derive or remember.Example 3. (A nonseparable inner-product space) . A normed linear space (or any topological space) is said to be separable if it contains a countable dense set. If an inner-product space is nonseparable, it cannot have a count
able orthonormal base. For an example, we consider the uncountable family of functions
u>. (t) = ei>'t,
wheret
E IR and ,\ E IR. This family of functions is linearly independent (Problem5
),
and is therefore a Hamel basis for a linear space X. We introduce an inner product in X by defining the inner product of two elements in the Hamel base:This is the value that arises in the following integration:
1
j
T - 1j
T .lim
2T u>.(t) uq(t) dt
== lim2T e,(>,-q)t dt
T_� -T T-� _T
76 Chapter 2 Hilbert Spaces
If A
=
a, this calculation produces the result1.
If A i-a , we getO.
Elements of X have the property of almost periodicity. (See Problem1.)
• Example 4. (Other abstract Hilbert spaces). A higher level of abstraction can be used to generate further inner product spaces and Hilbert spaces. Let us create at one stroke a Hilbert space of any given dimension. Let5
be any set.The notation
res
denotes the family of all functions from5
to the field IC. This set of functions has a natural linear structure, for ifx
andy
belong tores, x
+y
can be defined by
(x
+y)(s) = x(s)
+y(s)
A similar equation defines AX for ,\ E IC. Within
res
we single out the subspace X of allx
Eres
such that( 1) 2: [ lx(sW : s
E5 ] < 00
(Here we are using the notion of unordered sum as defined previously.) This . construction is familiar in certain cases. For example, if
S =
{ I ,2, .
. . , n}
, then the space X just constructed is the familiar spaceren.
On the other hand, ifS =
N, then X is the familiar space £2 . In the space X, addition and scalar multiplication are already defined, since X CreS.
Naturally, we define the inner product by(2) (x, y)
=2: [x(s)y(s) : s
ES]
r..luch of what we are doing here loses its mystery when we recall (from the Corollary to Theorem
4)
that the sums in Equations(1)
and(2)
are always countable. The space discussed here is denoted by £2(S). • Example 5. (Legendre polynomials.) An important example of an orthonormal basis is provided by the Legendre polynomials. We consider the space
C! - 1 , 1 J and use the simple inner product
( J
,g ) = [II f(t)g(t) dt
Now apply the Gram-Schmidt process to the monomials
t
o-t1 , t, t2, t3,
. • . The un-normalized polynomials that result can be described recursively, using the classical notationPn:
Po(t) =
12n - 1 n
- 1
Pn(t)
=--tPn_l (t) - --Pn-2(t)
n n
(
n= 2, 3, . . . )
The orthonormal system is, of course,