HILBERT SPACES

FRANZ LUEF

1. Orthogonality
Let M be a subspace of a Hilbert space H. Then the orthogonal complement of
M is defined by
M ⊥ = {x ∈ H : hx, yi = 0 for all y ∈ M }.
The linearity and the continuity of the inner product allow us to show the following
fact.
Lemma 1.1. M ⊥ is a closed subspace.
Proof. Let x be an element of the closure of M ⊥ . Hence there is a sequence (xn ) in M ⊥
such that xn → x. Now, for any y ∈ M ⊥ we have hx, yi = hlimn xn , yi = limn hxn , yi =
0. Hence x ∈ M ⊥ . 
Some elementary properties are
(1) M ⊂ N implies that N ⊥ ⊂ M ⊥ .
⊥
(2) M ⊆ M ⊥ .
A theorem of utmost importance is the closest point theorem for closed convex sets in
Hilbert spaces.
Theorem 1.2. If M is a closed convex set and x is any point of a Hilbert space H,
then there is a unique point x0 ∈ M which is closest to x: For all x ∈ H we have
that
kx − x0 k ≤ kx − yk for all y ∈ M.
This is not true in Banach spaces. For example, c0 is a closed subspace in `∞ , but
there is no closest sequence in c0 to the sequence (1, 1, 1, ...). In fact, the distance
between c0 and (1, 1, 1, ...) is 1, and this is achieved by any bounded sequence (an ) with
an ∈ [0, 2].

Important application: Let M be a closed subspace of H. In this situation, we are

able to express the least distance property in a geometrical manner.
Theorem 1.3 (Projection theorem). For a closed subspace M of H the point x0 ∈ M
is the closest point to an element x ∈ H if and only if x − x0 ∈ M ⊥ .
Corollary 1.4. H = M ⊕ M ⊥ .
Date: 10.09.2015.

1
 ⊥
Corollary 1.5. If M is a closed linear subspace of H, then M ⊥ = M . If S is a set
⊥
in H, then S ⊥ = span(S).
⊥
Proof. Let x ∈ M ⊥ . Then x = x0 + y0 where x0 ∈ M and y0 ∈ M ⊥ . We have
0 = hx, y0 i = hx0 , y0 i + hy0 , y0 i = ky0 k2 . Hence y0 = 0 and consequently, x ∈ M . 
⊥
Corollary 1.6. If S is a set in H, then S ⊥ = span(S). In particular, span(S) is
dense in H if and only if S ⊥ = 0.
Proof. Note that span(S) is the smallest closed subspace containing S. Therefore,
⊥
since S ⊆ span(S), we get by the preceding corollary that span(S) ⊆ S ⊥ and hence
⊥ ⊥⊥
S ⊥ ⊆ span(S) = span(S). 
We look at the mapping PM that takes a x ∈ H to its closest point x0 ∈ M .
Theorem 1.7. If M is a closed subspace of H, then the map P : H → H defined by
P x = x0 is a continuous projection with im(P ) = M, ker(P ) = M ⊥ and ker(P ) ⊥
im(P ).
Proof. By definition of P , for any x ∈ H we have x − P x ∈ M ⊥ and so P x ∈ M .

The mapping P is linear: We have x + y − (P x + P y) = x − P x + y − P y ∈ M ⊥

and P x + P y ∈ M . Hence P (x + y) = P x + P y. We also have P (αx) = αP x.

P is a projection: P 2 = P . Since P x ∈ M we have P 2 x = P x.

P is onto M : Since for any x ∈ M we have P x = x.

Moreover, x ∈ ker(P ) if and only if P x = 0 if and only if x = x − P x ∈ M ⊥ .

P is continuous: Since kxk2 = kx − P xk2 + kP xk2 by Pythagoras. Hence kP xk ≤

kxk. 
Look at the following example: Let M be the subspace of H spanned by an orthonor-
mal vector y. Then the orthogonal projection is given by P x = hx, yiy.

Another example: Find the quadratic polynomial closest to a function f in L2 [0, 1].
The space of quadratic polynomials P2 is three-dimensional and {1, x, x2 } is a basis
and P2 is a closed subspace of L2 [0, 1]. Hence by our general result, there exists a
p ∈ P2 closest to f . In fact, p is characterized by the condition p − f ⊥ q for all q ∈ M .
Equivalently, hp, xi i = hf, xi i for i = 0, 1, 2. The right hand sides can be computed,
because f is a given function. Hence we get three linear equations in three unknowns,
which can be solved.
2. Orthonormal Bases
A set of elements {en : n ∈ N} in H is called orthonormal if hei , ej i = δij . We call
it a basis if span{ei : ı ∈ N} is dense in H.

2
 (1) `2 has a basis consisting of {ei : i = 1, 2, ..} for ei = (0, 0, 0, ..., 0, 1, 0, ...). Since
spanei : i ∈ N is dense in `2 . Let x = (xi ) in `2 be orthogonal to all ei . Then
for any i we have xi = hx, ei i = 0 and hence x = 0.
P
If {en } is a basis for a Hilbert space, then we want to know when k ak ek converges.
Proposition 2.1. Let M be a closed subspace of H with a countable Hilbert basis en .
Then ∞ 2
P
k=1 k k converges in M if and only if (ak ) is in ` .
a e
Proof. Let xn = nk=1 ak ek . Then assuming n > m:
P

kxn − xm k2 = hxn − xm , xn − xm i
Xn m
X
=h ak ek , ak ek i
k=1 k=1
Xn n
X
=h ak ek , ak e k i
k=m+1 k=m+1
Xn n
X
= aj ak hej , ek i
j=m+1 k=m+1
Xn
= |aj |2 .
j=m+1

Suppose (ak ) ∈ `2 . Then this last sum converges to 0, implying that (xn ) is a Cauchy
sequence in M , which must converge to a point x ∈ M .

Conversely, suppose that the sequence (xn ) converges in M . Then it is a Cauchy

sequence, and kxn − xm k → 0, implying that ( j = 1 |aj |2 ) is a Cauchy sequence in
n
P
C and must converge as n → ∞. 
Theorem 2.2. If a closed subspace M of H has a basis (ek ), then we have for x ∈ H:
P∞ 2 2
(1) BesselPinequality: k=1 |hx, ek i| ≤ kxk ;
(2) P x = k hx, ek iek .
Proof. Let xn = nk=1 hx, ek iek . Then xn ∈ M and letting ak = hx, ek i we get
P

0 ≤ kx − xn k2 = hx − xn , x − xn i
= hx, xh−hxn , xi − hx, xn i + hxn , xn i
Xn Xn
2 2
= kxk − 2 |ak | + aj ak hej , ek i
k=1 j,k=1
n
X
= kxk2 − |ak |2 .
k=1

3
Hence, nk=1 |hx, ek i|2 ≤ kxk2 , which is an increasing sequence bounded above. Hence
P
the sum on the left must converge and this implies the Bessel inequality.
Pn
By the previous proposition the series k=1 hx, ek iek converges to a point y ∈ M .
Moreover for all m we have
n
X
hx − y, em i = hx, em i − hx, ek ihek , em i = 0.
k=1

Hence x − y ∈ M ⊥ and so y is the closest point in M to x. 

An important consequence of the previous theorem is the following result.

P identity). If (ek ) is aPbasis for2 a Hilbert space H, then for

Theorem 2.3 (Parseval’s
all x ∈ X we have x = k hx, ek iek and kxk2 = ∞ k=1 |ak | .

Proof. Since P x = x for M = H we get the basis expansion for all x ∈ H, and the
norm relation follows by the linearity and continuity of the innerproduct. 
Another consequence is that every separable Hilbert space is isomorphic to `2 .
Theorem 2.4. Consider the map J : H → `2 , x 7→ (hx, ek i). Then Bessel’s inequality
shows that the Fourier coefficients are in `2 . Parseval’s identity also implies that the
innerproducts are preserved: hJx, Jyi = hx, yi and so PJ is an isometry. Hence it is
injective, the map J is also surjective: Since it maps k hx, ek iek to (hx, ek i), which is
a well-defined element in `2 .

Department of Mathematics, Norwegian University of Science and Technology,

7491 Trondheim.
E-mail address: [email protected]