Hilbert
Hilbert
Hilbert space is a vector space with some extra structure. We start with
formal (axiomatic) definition of a vector space.
¦ For each two vectors |ai, |bi ∈ ν there exists a summation procedure:
|ai + |bi = |ci, where |ci ∈ ν. The summation obeys the following laws.
[Here we start using the symbol ∀ that means ‘for all’ (‘for each’, ‘for any’)
and the symbol ∃ that means ‘there exists’.]
1
(−1) · | ai = |−ai . (10)
| ai − | bi ≡ | ai + |−bi . (11)
Here the bar denotes complex conjugation. From Eq. (12) it directly follows
that the inner product of a vector with itself is always real. The next axiom
of inner product requires also that ha|ai be positive if |ai 6= |0i, and that
h 0 | 0 i = 0. Finally, we require that
h a | α u + β v i = α h a | ui + β h a | vi . (13)
h α u + β v | a i = α∗ h u | ai + β ∗ h v | ai . (14)
2
Linear self-adjoint operator. Linear operator, L, is a function from ν to
ν—that is it transforms one vector of the given space into another vector of
the same space—that satisfies the following requirement:
h b | L | a i = h a | L† | b i . (16)
h b | La i = h L† b | a i . (17)
Clearly, (L† )† ≡ L.
An operator L is self-adjoint (Hermitian) if L† = L, that is if ∀ |ai, |bi :
h b | La i = h Lb | a i . (18)
3
Comparing this to (20) we have
What if λu1 = λu2 ? In this case |u1 i and |u2 i are not necessarily orthogonal.
What is important, however, is that all the eigenvectors of the same eigen-
value λ form a vector sub-space of the original vector space, because any
linear combination of two eigenvectors with one and the same eigenvalue λ
makes another eigenvector with the same eigenvalue λ. Within this subspace
one can choose an orthogonal basis. This way we arrive at an orthogonal
basis of the eigenvectors of the self-adjoint operator, which is very useful for
many practical applications.
where {cj } are some complex numbers, is called a linear combination of the
vectors {|φj i}. As is directly seen from the axioms of the vector space, all
the linear combinations of the given set {|φj i} form some vector space ν̃—for
such a space we will be using the term sub-space to stress the fact that ν̃ ⊆ ν.
If ν̃ = ν, that is if any vector of the space ν can be represented as a linear
combination of the vectors {|φj i}, then the set {|φj i} is called basis and
the space ν is a finite-dimensional space. For any finite-dimensional vector
space there is a minimal possible number n for a vector set to form a basis.
This number is called dimensionality of the space, and, correspondingly, the
space is called n-dimensional space. In an n-dimensional vector space, all
the vectors of an n-vector basis {|φj i} are linear independent in the sense
that any linear combination of these vectors is different from zero if at least
one of the numbers cj is non-zero. Indeed, if the vectors {|φj i} were linear
dependent, that is if cj 6= 0 for at least one j = j0 , then the vector |φj0 i could
be expressed as a linear combination of the other vectors of the basis, and we
would get a basis of (n − 1) vectors—all the vectors of the original basis, but
the vector |φj0 i. But this is impossible by definition of the dimensionality.
4
Hence, without loss of generality we may deal only with a linear-indepen-
dent basis, since we can always eliminate linear dependent vectors. Below we
assume that our basis is linear independent, so that n is the dimensionality
of our space.
Given an arbitrary basis{|φj i}, we can construct an orthonormal basis
{|ej i}:
h ei | ej i = δij . (26)
[The term orthonormal means orthogonal and normal, where orthogonal
means h ei | ej i = 0, ∀i 6= j, and normal means h ei | ei i = 1, ∀i.] This
is done by the Gram-Schmidt orthonormalization procedure:
q
|ẽ1 i = |φ1 i , |e1 i = |ẽ1 i/ hẽ1 |ẽ1 i ,
q
|ẽ2 i = |φ2 i − he1 |φ2 i |e1 i , |e2 i = |ẽ2 i/ hẽ2 |ẽ2 i ,
q
|ẽ3 i = |φ3 i − he1 |φ3 i |e1 i − he2 |φ3 i |e2 i , |e3 i = |ẽ3 i/ hẽ3 |ẽ3 i ,
· · · · · · · · · · · · · · · · · · · ·q · · ·
|ẽn i = |φn i − he1 |φn i |e1 i − . . . − hen−1 |φn i |en−1 i , |en i = |ẽn i/ hẽn |ẽn i .
α (x1 , x2 , . . . , xn ) = (α x1 , α x2 , . . . , α xn ) , (28)
(x1 , x2 , . . . , xn ) · (y1 , y2 , . . . , yn ) = x∗1 y1 + x∗2 y2 + . . . + x∗n yn . (29)
Fixing some orthonormal basis {|ej i} in our vector space ν, we first note
that ∀ |xi ∈ ν, the coefficients xj —referred to as components of the vector
|xi with respect to the given basis—in the expansion
n
X
|xi = xj |ej i , (30)
j=1
5
which is seen by constructing inner products of the r.h.s. of (30) with the
basis vectors. This leads to a one-to-one mapping between the vectors of
the spaces ν and C n :
|xi ↔ (x1 , x2 , . . . , xn ) . (32)
By the axioms of the inner product, from (30)-(31) one makes sure that if
|xi ↔ (x1 , x2 , . . . , xn ) and |yi ↔ (y1 , y2 , . . . , yn ), then
|xi + |yi ↔ (x1 , x2 , . . . , xn ) + (y1 , y2 , . . . , yn ) , (33)
α |xi ↔ α (x1 , x2 , . . . , xn ) , (34)
hx|yi = (x1 , x2 , . . . , xn ) · (y1 , y2 , . . . , yn ) , (35)
That is we see a complete equivalence of the two vector spaces, and it is
precisely this type of equivalence which we understand by the word isomor-
phism.
Problem 22. Consider the vector space R2 , i.e. the set of pairs (x, y) of real
numbers.
(a) Show that the function k · kM defined by k(x, y)kM = max{|x|, |y|} is a norm
on R2 .
(b) Show that the function k · kS defined by k(x, y)kS = |x| + |y| is a norm on R2 .
(c) In any normed space ν the unit ball B1 is defined to be {u ∈ ν | kuk ≤ 1}. Draw
the unit ball in R2 for each of the norms, k · kM and k · kS .
6
Problem 23. Prove that with the definition (38): (i) all the axioms of norm are
satisfied and (ii) there take place the following specific to the inner-product norm
properties.
hy|xi = 0 ⇒ kx + yk2 = kxk2 + kyk2 , (39)
Cauchy-Bunyakovsky-Schwarz inequality:
moreover,
|h y | x i| ≤ kxk · kyk , (41)
parallelogram law:
Hints. The triangle inequality directly follows from the simplest version of Cauchy-
Bunyakovsky inequality, Eq. (40). To prove Eq. (40), utilize the fact that the
product h x + λy | x + λy i is a second order polynomial of λ which is non-negative
∀λ (by an axiom of the inner product), which implies a certain constraint on its
discriminant. To arrive at Eq. (41), use the same approach, but with λ → λh y | x i.
The question now is: Does any Cauchy sequence in a given vector space
ν converge to some vector x ∈ ν? The answer is not necessarily positive
and essentially depends on the structure of the vector space ν. (Normed
7
spaces where all Cauchy sequences converge are called complete spaces, or
Banach spaces.) For any inner-product vector space of a finite dimension
the answer is positive and is readily proven by utilizing the fact of existence
of the orthonormal basis, {|ei i}. First we note that if
n
X
|ai = α i | ei i , (44)
i=1
then q
kak = |α1 |2 + |α2 |2 + . . . + |αn |2 . (45)
If we have a Cauchy sequence of vectors |a(k) i, then for any given i =
(k)
1, 2, · · · , n the i-th coordinates of the vectors, αi , form a Cauchy sequence
of complex numbers. Any Cauchy sequence of complex numbers converges
to some complex number.—This is a consequence of the fact that a complex-
number Cauchy sequence is equivalent to two real-number Cauchy sequences
(for the real and imaginary parts, respectively) and a well-known fact of the
theory of real numbers that any real-number Cauchy sequence is convergent
(completeness of the set of real numbers). We thus introduce the numbers
(k)
αi = lim αi , (46)
k→∞
and easily see that our vector sequence converges to the vector
n
X
|ai = αi | ei i . (47)
i=1
Problem 25. Show by an example that the ‘sup’ norm does not imply the paral-
lelogram law.
8
Hence here we have an example of a Banach space which is not a Hilbert
space.
is called Fourier series for the vector |xi with respect to the given orthonor-
mal systems; the numbers hej |xi are called Fourier coefficients.
Theorem. Partial sums of the Fourier series form a Cauchy sequence.
Proof. We need to show that {|x(n) i}, where
n
X
(n)
|x i= hej |xi |ej i (49)
j=1
Hence, we obtain
n
X
|hej |xi|2 ≤ kxk2 (Bessel inequality) , (54)
j=1
9
and prove the theorem.
Moreover, rewriting Eq. (53) as
v
u n
X
(n) u
kx − x k = tkxk2 − |hej |xi|2 . (55)
j=1
we see that for the series (49) to converge to the vector |xi it is necessary
and sufficient to satisfy the condition
∞
X
|hej |xi|2 = kxk2 (Parseval relation) . (56)
j=1
10
Completion. Actually, any normed vector space can be upgraded to a
complete space. And any inner-product vector space can be upgraded to a
Hilbert space. Corresponding vector space is called completion. The proce-
dure of completing an incomplete vector space ν is as follows. Consider a
set ν̃ of all Cauchy sequences {xk } ∈ ν. The set ν̃ is a vector space with
respect to the addition and multiplication by a complex number α defined
as
| {xk } i + | {yk } i = | {xk + yk } i , (57)
α | {xk } i = | {α xk } i . (58)
[That is the sum of sequences is defined as the sequence of sums; the product
of a sequence and a number is the sequence of products.] We also need to
introduce the equality as
11
The construction of the isomorphism is absolutely the same as in the pre-
viously considered case of the finite-dimensional space, and we do not need
to repeat it. The only necessary remark is that the convergence of corre-
sponding series in the space l2 is guaranteed by Eq. (62). This isomorphism
means that all the Hilbert spaces with a countably infinite basis have a
similar structure which is very close to that of a finite-dimensional Hilbert
space.
and it is easy to show that the space C[a, b] is not complete with respect
to this norm. Indeed, consider a Cauchy sequence of functions fk (x), x ∈
[−1, 1], where fk (x) = 0 at x ∈ [−1, 0], fk (x) = kx at x ∈ [0, 1/k], and
fk (x) = 1 at x ∈ [1/k, 1]. It is easily seen that in the limit of k → ∞
the sequence {fk } converges to the function f (x) such that f (x) = 0 at
x ∈ [−1, 0], f (x) = 1 at x ∈ (−1, 0]. But the function f (x) is discontinuous
at x = 0, that is f ∈/ C[a, b].
Is it possible to construct a Hilbert space of functions? The answer is
“yes”, but the theory becomes quite non-trivial. Here we confine ourselves
with outlining some most important results. First, one has to understand
the integral in the definition of the inner product (63) as a Lebesgue integral,
which is a more powerful integration scheme than the usual one. Then, one
introduces the set of functions, L2 [a, b], by the following rule: f ∈ L2 [a, b] if
the Lebesgue integral
Z b
|f |2 dx (65)
a
does exist. For the set L2 [a, b] to be a vector space, one has to introduce a
weaker than usual notion of equality. Namely, two functions, f (x) and g(x),
12
if considered as vectors of the space L2 [a, b], are declared “equal” if
Z b
|f − g|2 dx = 0 . (66)
a
which converges to |f i by the inner product norm, that is the series (67)
converges to f (x) almost everywhere. The basis of polynomials {xk , k =
0, 1, 2, . . .} is not orthogonal. However, one can easily orthonormalize it
by the Gram-Schmidt process. The orthonormalized polynomials are called
Legendre polynomials. Clearly, the particular form of Legendre polynomials
depends on the interval [a, b].
It is worth noting that while the space L2 [a, b] contains pretty weird
functions—for example, featuring an infinite number of jumps—the poly-
nomial basis consists of analytic functions. This is because any function in
L2 [a, b] can be approximated to any given accuracy—in the integral sense
of Eq. (63)—by some polynomial. In this connection they say that the set
of all polynomials is dense in L2 [a, b].
13