You are on page 1of 37

Quantization of Analysis

Kelvin Lui
12/11/2014

Department of Mathematics
Colby College
2015

Definitions

Definition 1. A C -algebra, A, is a Banach algebra over the field of complex numbers with a map , such that
x = (x ) = x
(x + y) = x + y
(xy) = y x

(x) = x
kx xk = kxk2

(1.0.1a)
(1.0.1b)
(1.0.1c)
(1.0.1d)
(1.0.1e)

for all x, y in A and in C.


Definition 2. An element of a C - algebra, A, is positive if and only if it is selfadjoint and its spectrum is contained in the nonnegative reals, or equivalently, if it is
of the form a a from some element a in A.
Definition 3. A is a positive operator if the inner product hAx, xi > 0 for every
vector x in A.
Definition 4. A C - algebra, A, is unital if it admits a multiplicative identity 1
such that
1 a = (a 1) = a
a1 = a
k1k = k1 1| = k1k2

(1.0.2a)
(1.0.2b)
(1.0.2c)

Definition 5. If S is an operator system, B is a C -algebra, and : S B is a


linear map, then is called a positive map provided that it maps positive elements
of S to positive elements of B.
Definition 6. If B is a C -algebra and : S B is a linear map, where S is an
operator system, then we define n : Mn (S) Mn (B) by n ((ai,j )) = ((ai,j )). We
call completely positive if is n-positive for all n.
Definition 7. A linear algebra homomorphism between C - algebras : A B
which is self-adjoint. ie. (a ) = (a) is called a -homomorphism. A unital homomorphism is such that (1A ) = (1B ).

2
2.1

Abstract Algebra
Rings and Ideals

Rings are essential tools to understand... blah blah blah

Definition 8. A ring is a datum (R, +, , 0, 1) where R is a set, 1, 0 R and +,


are binary operations on R defined such that
1. R is an abelian group under + with identity element 0;
2. The binary operator is associative and 1 x = x 1 = x x R;
3. Multiplication distributes over addition:
x (y + z) = (x y) + (x z)
(x + y) z = (x z) + (y z) x, y, z R
If the operation is commutative, x y = y x x, y R, we say R is a
commutative ring. In addition, if 1 = 0 then the ring has only one element and is
referred to as the zero ring. An example of a ring would be the integers Z or the
integers modulo n Z/nZ , n Z. If R is a ring, we can construct a new ring R[t] of
polynomials in t with coefficients in R.
Definition 9. If R is a ring, a subset S R is said to be a subring if it inherits
the structure of a ring from R, therefore 0, 1 S and S is closed under the addition
and multiplication operations in R; (S, +) is a subgroup of (R, +).
An example would be the integers Z are a subring of Q.
Lemma 2.1. Subring Criterion
Let R be a ring and S R, then S is a sunning if and only if 1 S and s1 , s2 S
we have s1 s2 , s1 s2 S.
Proof. () If we assume that S is a subring the proof is trivial.
() Note S 6= {0} because 1 S. If s1 s2 , s1 s2 S, s1 , s2 S that implies
that S is an additive subgroup by the subgroup test. It is also easy to see that 0 S
and that the other conditions for a subring hold.
Definition 10. A map f : R S between rings R and S is said to be a (ring)
homomorphism if
1. f (1R ) = 1S
2. f (r1 +R r2 ) = f (r1 ) +S f (r2 )
3. f (r1 .R r2 ) = f (r1 ).S f (r2 )

The image of a ring homomorphism f : R S is a subring of S. If the image is


all of S we say f is surjective, and if f : R S is an isomorphism then g : S toR
such that f g = idS and g f = idR .
We can also look at polynomials with coefficients in a ring. If R is a ring then the
set R[t] of sequences (an )nN , an R where N N such that an = 0, n N . We
can denote elements of R[t] as
N
X

an tn .

(2.1.2)

n=0

We will define addition and multiplication to be


N
X
n=0
N
X
n=0

an t +

an tn

M
X
n=0
M
X
n=0

max{N,M }
n

bn t =
bn tn =

(an + bn )tn

n=0
NX
+M X
n
n=0


ak bnk tn

(2.1.3a)

(2.1.3b)

k=0

and with this we see that R[t] is a ring. The set of all sequences (an )nN forms a
ring as well with the same definitions of addition and multiplication. It is called the
ring of formal power
P seriesn in t and denoted as R[[t]]. We view elements of R[[t]] as
infinite sums , n+0 an t .
Definition 11. If R is a ring, then an element a R\{0} is a zero-divisor if there
is some b R\{0} such that a.b = 0.
Definition 12. Let f : R S be a ring homomorphism. The kernel of f is
ker(F ) = {r R : f (r) = 0},

(2.1.4)

im(f ) = {s S : r R, f (r) = 2}.

(2.1.5)

and the image of f is

The image of a homomorphism is a subring of the target ring. The kernel, however
is not a subring of the target ring. The kernel is closed under addition and multiplication and since 0.x = 0, x, it obeys a stronger kind of closure with respect to
multiplication. If x ker(f ) and r R then f (x.r) = f (x).f (r) = 0.f (r) = 0 so
that x.r ker(f ).
Definition 13. Let R be a ring. A subset I R is called an ideal if it is a subgroup
of (R, +) and moreover for any a I and r R then a.r I.
Lemma 2.2. If f : R S is a homomorphism, then ker(f ) is an ideal.

Proof. Evident from the given definitions.


Lemma 2.3. Let R be a ring, and I, J ideals in R. Then I + J, I J, and IJ are
ideals, where

I + J = {i + j : i I, j J};

IJ = {

n
X

ik jk : ik I, jk J, n N }.

(2.1.6)

k=1

Moreover we have it that IJ I J, and I, J I + J.


Proof. For both I + J and I J it is easy to see and prove that both are ideals in R.
To see that IJ is an ideal, note that the sum of two elements in IJ is
n
X

ik jk +

k=1

m
X

max(n,m) 

pk qk =

k=1

ik jk + pk qk




k=1

ik , pk I, jk , qk J, n, m N,
P
which is of the same form. If nk=1 xk yk IJ then

n
X
k=1

x k yk =

n
X

(xk )yk IJ,

(2.2.4)

(2.1.8)

k=1

because if xk I xk I. IJ is an abelian subgroup. It is clear to see that


i, J I + J and IJ I J:
I = I + 0 I + J;

J =0+J I +J

(2.1.9)

and let i I, j J so that ij IJ then ij I J because I and J are ideals.


Finally, I J is closed under addition so every sum of products of the form ij is also
in I J.

2.2

Quotients

We will show later that ideals and kernels of ring homomorphisms are the same,
however we first need to know the notion of quotient for rings.
Let us suppose that R is a ring and I is an ideal in R. Since (I, +) is a subgroup of
the abelian group (R, +), we may form the quotient group (R/I, +); the construction
of the group being: r s if r s I, and the equivalence classes are the cosets
{r + I : r R}. To endow R/I with the structure of an abelian group suppose that
C1 , C2 R/I then we define

C1 + C2 = {c1 + c2 : c1 Ci , i = 1, 2}

(2.2.1)

C1 + C2 is a subset of R and we claim it to be a single coset of I. Definitely, it


is a union of cosets, since c1 + c2 r, for some r R we have r = (c1 + c2 ) + i for
some i I and so r = (c1 + i) + c2 C1 + C2 ; the sum is an element of the union
of cosets. To show that it is a single coset, that addition is well-defined, fix elements
di Ci , i = 1, 2. Then if ci Ci we have ci = di + ki , ki I, hence
c1 + c2 + (d1 + k1 ) + (d2 + k2 ) = (d1 + d2 ) + (k1 + k2 ) (d1 + d2 ) + I.

(2.2.2)

To give R/I a ring structure, we need to define a multiplication operation. Given


two cosets C1 , C2 define:
C1 C2 = {c1 c2 : ci Ci , i = 1, 2}.

(2.2.3)

We also claim that the set C1 C2 is contained in a single coset of R/I, that is
C1 C2 + I is a single I-coset. Suppose the ci , di are given as above, then
c1 c2 = (d1 + k1 ) (d2 + k2 ) = d1 d2 + (d1 k2 + k1 22 + k1 k2 ) d1 d2 + I (2.2.4)
{z
}
|
I

We get a binary operation on R/I by setting:


C1 C2 = C1 C2 + I

(2.2.5)

Theorem 2.4. The datum (R/I, +, , 0 + I, i + I) defines a ring structure on R/I


and moreover the map 1 : R R/I given by q(r) = r + I is a surjective ring
homomorphism.
Proof. It is clear that R/I is an abelian group under + with an identity element
0. We then check to see that the binary operation is associative and that the
multiplication distributes over addition:
C1 (C2 C3 ) = (a1 + I) ((a2 + I) (a3 + I))
= (a1 + I)((a1 a2 + I)) = (a1 )(a2 a3 ) + I
= (a1 a2 )a3 + I = (a1 a2 + I) (a3 + I)
= ((a1 + I) (a2 + I)) (a3 + I)
= (C1 C2 ) C3
C1 , C2 , C3 R/I, a1 , a2 , a3 R

(2.2.6)

a simple adaptation with the multiplicative identity will show associativity of


1R/O C1 = C1 = C1 1R/O , where 1R/O = 1 + I.
C1 (C2 + C3 ) = (a1 + I) ((a2 + I) + (a3 + I))
= (a1 + I) ((a2 + a3 ) + I) = a1 (a2 + a3 ) + I
= (a1 a2 + a1 a + 3) + I
= (a1 a2 + I) + (a1 a3 + I)
= C1 C2 + C1 C3
C1 , C2 , C3 R/I, a1 , a2 , a3 R

(2.2.7)

A similar process can be done to show (C1 + C2 ) C3 = (C1 C3 ) + (C2 C3 ).


Finally the map q : R R/I is surjective and that it is a homomorphism can
been seen immediately from the definitions.
The map q : R R/I is called the quotient homomorphism.
Corollary 2.4.1. Any ideal is the kernel of a ring homomorphism.
Proof. If I is an ideal and q : R R/I is the quotient map then clearly ker(q) = I.
Next, we will look at comparisons between ideals in a quotient ring with ideals in
the original ring.
Lemma 2.5. Let R be a ring, I an ideal in R and q : R R/I the quotient map.
If J is an ideal then q(J) is an ideal in R/I, and if K is an ideal in R/I then
q 1 (K) = {r R : q(r) K} is an ideal in R which contains I. Moreover, these
correspondences give a bijection between the ideals in R/I and the ideals in R which
contain I.
Proof. Let q : R R/I denote the quotient homomorphism. If J is an ideal in
R we need to check that q(J) is an ideal in R/I. q is a homomorphism so it is a
subgroup of R/I under addition, we need to check the multiplicative property. Any
element of q(J) has the form q(j), j J. From Theorem 2.4 we know that the
map q is surjective and so for any element of R/I has the form q(r), r R. Since
q(r) q(j) = q(r j) and r j Jj then q(J) is an ideal in R/I.
If K R/I is an ideal then q 1 (K) is an ideal of R. Let t1 , t2 q 1 (K) the
q(t1 + t2 ) = q(t1 ) + q(t2 ) K because K is an ideal, then t1 + t2 q 1 (K). If
= q(r) q(t) K because K is an ideal of R/I.
t q 1 (K) and r R then q(rt)
1
Then this means that r t q (K) and therefore q 1 (K) is an ideal in R.
Note that q(q 1 (K)) = K, true for surjective map q of sets. If J is an ideal of R
then we claim that the composition q 1 (q(J)) is the ideal J + I. Assuming that if

I J then q 1 (q(J)) = J, hence q is bijective when restricted to ideals containing I


which is required. Now we need to check that q 1 (q(J)) + J + I. Let x q 1 (q(J))
then q(x) = q(j) for some j J and so q(x j) = 0, then x j I, because
q(J) R/I. However, x = j + (x j) J + I so that q 1 (q(J)) J + I. The reverse
inclusion is clear because q(j + i) = (j + i) + I = j + I, j J, i I.
Theorem 2.6. First Isomorphism theorem
Let f : R S be a homomorphism of rings, and let I = ker(f ). Then f induces
an isomorphism f : R/I im(f ) given by:
f(r + I) = f (r)

(2.2.8)

Proof. Note that if r s I then f (r s) = 0 and therefore f (r) = f (s), f take a


single value on each coset r + I so f is well-defined. Clearly from the definition of the
ring structure on R/I it is a ring homomorphism. It is clearly surjective, if s im(f )
then s = f (r) for some r R thus s = f(r + I). To see that f is injective it is enough
that ker(f) = 0. If f(R + I) = 0 f (r) = 0 hence r I so that r + I = I.
Let R be a ring and A, B be subsets of R. Let A + B = {a + b : a A, b B}.
Theorem 2.7. Second Isomorphism theorem
Let R be a ring and A a subring of R, B an ideal of R. Then A + B is a subring of
R and the restriction of the quotient map q : R R/B to A induces an isomorphism
A/A B
= (A + B)/B

(2.2.9)

Proof. First let us check that A + B is a subring of R. It is clear that 1 A + B


since 1 A and for 1 A + B, 1 = 1 + 0. Then using the two other conditions of the
subring criterion
(a1 + b1 ) (a2 + b2 ) = a1 a2 + b1 a2 + a1 b2 + b1 b2 A + B
{z
}
|

(2.2.10a)

(a1 + b1 ) + (a2 + b2 ) = (a1 + a2 ) + (b1 + b2 ) A + B

(2.2.10b)

It is easy to see that A B is an ideal in A and that B is an ideal in A + B, so


therefore the two quotients A/A B and (A + B)/B exist. We let q : R R/B
be the quotient map. The restriction of q to p : A R/B, whose image is clearly
(A + B)/B, so by the first isomorphism theorem it is enough to check that the kernel
of p is in A B. It is clear: if a A has p(a) = 0 then a + B = 0 + B so that a B
and so a A B.

Theorem 2.8. Universal property of Quotients


If f : R S is a homomorphism and I ker(f ) is an ideal, then there is a
unique homomorphism f : R/I S such that f = f q where q : R R/I is the
quotient map.
Proof. Note that since q is surjective the requirement that f(q(r)) = f (r) uniquely
determines f if it exists. But if r1 r2 I then since I ker(f ), we have 0 =
f (r1 r2 ) = f (r1 ) f (r2 ) hence f (r1 ) = f (r2 ). Follows that f is constant on the Icosets and so it induces a unique map f : R/I S such that fq = f . It is immediate
from the definitions of the ring structure on R/I that f is a homomorphism.
Theorem 2.9. Third Isomorphism theorem
If I J are ideals, then J/I = {j + I : j J} is an ideal in R/I and
(R/I)/(J/I)
= R/J

(2.2.11)

Proof. Let qI : R R/I and let qJ : R R/J be the two quotient maps. By the
universal property for the quotient qJ : R toR/J/ applied to the homomorphism qI we
see that there is a homomorphism qI : R/J R/I induced by the map qI : R R/I
and q qJ = qI . Clearly qI is surjective since qI , and if q(r + j) = 0 then since
r + J = qJ (r) so that qI (r + J) = qI (r) we have r I so that ker(
qI ) = J/I and the
result follows from the first homomorphism theorem.

2.3

Modules

Definition 14. Let R be a ring with identity 1R . A module over R is an abelian


group (M, +) together with a multiplication action of R on M satisfying:
1R .m = m, m M
(r1 .r2 )m = r1 (r2 .m), r1 , r2 R, m M
(r1 + r2 )m = r1 .m + r2 .m, r1 , r2 R, m M
r.(m1 + m2 ) = r.m1 + r.m2 , r R, m1 , m2 M

(2.3.1a)
(2.3.1b)
(2.3.1c)
(2.3.1d)

A concrete sense of a module M over a ring R, could be quickly understood if one


were to let R be a field, then the module would simply be a vector space over R. A
module over a field are essentially vector spaces over that field.
Definition 15. If M is an R-module, a subset N M is called a submodule if it
is an abelian subgroup of M and whenever r R and n N then r.n N .
One extension from linear algebra is that of linear independence.

Definition 16. If M is a module over R, we say a set S M is linear independent if


whenever we have an equation r1 s1 + r2 s2 + ... + rk sk = 0 for ri R, s1 S 1 i k
we have r1 = r2 = ... = rk = 0. We can say that a set S is a basis for a module M if
and only if it is linearly independent and it spans M . Any module which has a basis
is called a free module.
On the other hand there is a notion of a spanning set.
Definition 17. A spanning set for an R-module is a subset whose finite R-linear
combinations fill up the module. These always exist, since the entire module is a
spanning set.
To continue on we assume that all rings R are integral domains. As for vector
spaces, modules have a natural notion of a quotient module. An essential aspect of
the quotient module is the condition that N be a submodule so that the multiplication
on M induced a module structure on M/N .
Definition 18. If N is a submodule of M , then in particular it is a subgroup of an
abelian group, so we can form the quotient M/N . The module M/N is called the
quotient module of M by N . We define multiplication on the quote module by:
r.(m + N ) = r.m + N, rinR, m + N M/N

(2.3.2)

The multiplication well defined because if m1 + N = m2 + N we then have m1


m2 N , same as quotient rings, and so r.(m1 m2 ) N r.m1 + N = r.m2 + N .
Similar to rings, we can define a natural analogue of linear maps.
Definition 19. If M1 , M2 are modules, we say that : M1 M2 is a module
homomorphism if:
(m1 + m2 ) = (m1 ) + (m2 ), m1 , m2 M1
(r.m) = r.(m), r R, m M

(2.3.3a)
(2.3.3b)

Lemma 2.10. Submodule correspondence


Let M be an R-module and N a submodule. Let q : M M/N be the quotient
map. If S is a submodule of M then q(s) is a submodule of M/N , while if T is a
submodule of M/N then q 1 (T ) is a submodule of M . Moreover, the map T 7 q 1 (T )
gives an injective map from submodules of M/N to the submodules of M which contain
N , thus submodules of M/N correspond bijectively to submodules of M which contain
N.
Proof. We can check from the definitions that q(S) and q 1 (T ) are submodules of N
and M respectively. As the proof for both are the same, we will only prove it for
q(S).
Let S be a submodule of M and q : M M/N be the quotient map. Let
s1 , s2 S then q(s1 ), q(s2 ) M/N . Since S is a submodule then s1 + s2 S

q(s1 + s2 ) = q(s1 ) + q(s2 ) M/N . Finally, if r R then r.s S, by definition of


submodule, and (r.s) = r.(s) M/N . Therefore q(S) is a submodule of M/N .
Now, let T be any subset of M/N . We have that q(q 1 (T )) = T is surjective
because q is surjective. Also because know that q 1 (T ) is always a submodule in M ,
the map of S 7 q(S) is a surjective map from submodules in M to submodules in
M/N and that T 7 q 1 (T ) is an injective map, furthermore since q(N ) = 0 T
for any submodule T of M/N we have N q 1 (T ) so that the image of the map
T 7 q 1 (T ) consists of submodules of M which contain N . We only now need to
check that the submodules of M of the form q 1 (T ) are precisely these submodules.
Suppose that S is an arbitrary submodule of M and consider q 1 (q(S)). By
definition
q 1 (q(S)) = {m M : q(m) q(S)}
= {m M : s S  m + N = s + N }
= {m M : m s + N }

(2.3.4)

The right hand side is simply writing the submodule S + N , S/N . However, if
N S then we have S + N = S and therefore q 1 (q(S)) = S and any submodule S
which contains N is indeed the preimage of a submodule M/N .
Theorem 2.11. Universal property of quotients
Suppose that : M N is a homomorphism of R-modules, and S is a submodule
of M with S ker(). Then there is a unique homomorphism : M/S N such
where q : M toM/S is the quotient homomorphism, that is the following
that = q
diagram commutes:

M
q

M/S
is the submodule ker()/S = {m + S : m ker()}.
Moreover, ker()

Proof. q is surjective, so the formula (q(m))


= (m) determines uniquely the values
0

of , thus is unique if it exists. If mm S then because by assumption S ker()


it follows that 0 = (m m0 ) = (m) (m0 ) since is a homomorphism. Therefore
+ S) = (m). is then a
is constant on the S-cosets. This induces a map (m
homomorphism because it follows directly from the module structure on the quotient
M/S, and by definition = q.



(m1 + S) + (m2 + S) = m1 + m2 + S
= (m1 + m2 )
= (m1 ) + (m2 )
1 + S) + (m
2 + S)
= (m


r.(m + S) = (r.m
+ S)
= (r.m) = r.(m)
+ S)
= r.(m

(2.3.5a)

(2.3.5b)

+ S) = (m) = 0 if and only if m ker(), therefore the kernel of


Note that (m

is m + S ker()/S.
As we have done above, we will now look at the isomorphism theorems for modules.
First we let M be an R-module.
Theorem 2.12. First Isomorphism theorem
If : M N is a homomorphism then induces an isomorphism : M/ker()
im().
Proof. Let K = ker(). Using the universal property of quotients to K, then we have
= ker()/ker() = 0 it follows that is injective and therefore induces an
ker()
isomorphism onto its image which from the equation q = = Im().
Theorem 2.13. Second isomorphism theorem
If M is an R-module and N1 , N2 are submodules of M then
(N1 + N2 )/N2
= N1 /N1 N2

(2.3.6)

Proof. First, let q : M M/N2 be the quotient map, which restricts p from N1 to
M/N2 , and whose image is clearly (N1 + N2 )/N2 . Therefore by the first isomorphism
theorem it is only necessary to check that the kernel of p is N1 N2 . If n N1 has
p(n) = 0 then n + N2 = 0 + n2 so that m N2 and so n N1 N2 .
Theorem 2.14. Third isomorphism theorem
Suppose that N1 N2 are submodules of M . Then we have
(M/N1 )/(N2 /N1 )
= M/N2

(2.3.7)

Proof. We first let qi : M M/Ni for i = 1, 2. Using the universal property of


quotients for q1 we see that there is a homomorphism q2 : M/N1 M/N2 induced
by the map q2 : M M/N2 , with the kernel being ker(q2 /N1 ) = N2 /N1 and q2
q1 = q2 . Therefore q2 is surjective because q2 is and the result follows from the first
isomorphism theorem.
These proofs mainly followed that of the proofs for the isomorphism theorems of
rings.

3
3.1

Tensors and Tensor Products


Dual Spaces

The study of dual spaces is essential to understanding the notion of a tensor product.
To formalize our understanding of the dual space we will look at real-valued functions
defined on a real vector space T , f : T R. The set of all such functions is then
given the structure

(f + g)(v) = f (v) + (v),


(f )(v) = (f (v))
0(v) = 0
(f )(v) = (f (v))

v T
v T
v T
v T,

(3.1.1a)
(3.1.1b)
(3.1.1c)
(3.1.1d)

which defines a vector space. Furthermore, we will restrict the space of all realvalued functions to only those that are linear
f (u + v) = f (u) + f (v) , R, u, v T.

(3.1.2)

The real-valued linear functions on a real vector space are also called linear functionals and the space constitutes a vector space called the dual of T , denoted by
T .
For simplicity we will point out the two types of vectors, those in T an those in

T , and distinguish them from one another. For the basis vectors of T we will use
superscripts and the components of vectors we will use subscripts, therefore
if {ea }
P
is a basis of T , then a vector T has the unique expression = a a ea .
In a natural way, we will now prove that the that the dual space of T has the
same dimensionality as T .
Lemma 3.1. The dual space,T , of T has the same dimensionality as T .

Proof. First, let T be a N -dimenionsal vector space and let {ea } be the given basis
of T . We define {ea } to be real-valued functions which maps T into the real
number a , which is the ath component relative to the basis vectors of T , {ea }. They
are given to be
ea () = a T
ea (eb ) = ba

(3.1.3a)
(3.1.3b)

This clearly gives N real-valued functions which satisfy Eq(3.1.3b). Now we will
show that they are linear and they constitute a basis for T .
The linearity of the real-valued functions is shown by the restriction
ea ( + ) = a + a

, R , T

(3.1.4)

To prove that {ea } is a basis for any given T we can define N real numbers,
denoted a by (ea ) = a . Then
() = (a ea ) = a (ea )
= a a
= a ea () T

(2.1.5)

Therefore for any T we have = a ea , showing that {ea } stands T . The


independence of the basis vectors comes from Eq(3.1.3b). Thus dim(T ) = dim(T ).
Theorem 3.2. Let F be a linear functional on a normed space (E, k k). The
following statements are then equivalent to each other:
1. F is continuous;
2. F is continuous at 0;
3. sup {|F (x)| : x E, kxk 1} <
F is continuous if and only if it is bounded on the unit ball of E.
Proof. From statement 1 the proof of statement two is trivial.
Assuming statement 2, we know  > 0 > 0 such that kxk < |F (x)| < .
Take  = 1 then there exists a > 0 such that kxk < |F (x)| < 1. Therefore
k < |F ( x
)| < 1. Therefore |F (x)| < 2
x E such that kxk 1 we have k x
2
2
because of linearity.
Suppose statement 3 and let M be the finite supremum. Taking any pair of
(xy)
x, y E, kxyk
is a unit vector and therefore statement 3 holds. Then


(x y)
F
M
kx yk
|F (x) F (y)| M kx yk;
line 2 follows from linearity of the linear functional. Let
that F is continuous.

(2.1.6)

M

then it is clear

Lemma 3.3. Let E be the set of all continuous linear functionals on the normed
space (E, k k) and F E . Then kF k = supxE,kxk1 |F (x)| is a norm on E .
Proof. We simply need to check if the norm satisfies the following three conditions:
1. kF k > 0 if F 6= 0;
2. kF k = ||kF k C F E
3. kF + F 0 k kF k + kF 0 k F, F 0 E
Clearly conditions 1 and 2 are easily shown by definition of the norm. Condition
3 is satisfied by the triangle inequality.
Theorem 3.4. The set E of all continuous linear functionals on the normed space
(E, k k) is itself a Banach space with respect to pointless algebraic operations and
norm


kF k = supxE,kxk1 F (x)

(3.1.7)

Proof. E is a vector space over the same field as E. From Theorem 2.3, we know
that kF k is a real number and that it is finite. Lemma 2.3 also tells us that the given
norm is a norm on E . We will now show that E is complete.
Let (Fn ) be a Cauchy sequence in E so that kFn Fm k 0 as n, m .
Therefore x E |Fn (x) |Fm | 0 as n, m . We see that this is Cauchy
sequence of scalars; we denote the limit of this scalar sequence by F (x). We must
now show that F E and that Fn F with respect to the norm on E .
Let  > 0 and pick n0 N such that m, n n0 kFm Fn k < 
x E such that kxk 1 and m, n n0 ,
|Fm (x) Fn (x)| < 

(3.1.8)

Letting m , n n0 and x  kxk 1,


|F (x) Fn (x)| 

(3.1.9)

Since x E  kxk 1 implies F F0 and Fn0 are bounded, then F is also


bounded on the unit ball and by Theorem 2.3 it is continuous. Thus F E .
Whenever n n0 kF Fn k  implies that Fn F E and so E is complete.
Theorem 3.5. For any vector x in a normed space E and any continuous linear
functional F on E,
|F (x)| kF kkxk.
Proof. Suppose that x E  x 6= 0. Then

x
kxk

(3.1.10)

is a unit vector and therefore


x
F
kF k
kxk
|F (x)|
kF k
kxk

(2.1.11)

Lemma 3.6. Let V be an inner product space. x, y, z V , if (x, z) = (y, z)


V x = y.

Proof. If (x, z) = (y, z) then


0 = (x, z) + (1)(y, z)
= (x, z) + (y, z)
(x y, z)

(2.1.12)

If this is true for all z V then z = x y x y = 0.


Lemma 3.7. Let M be a linear subspace of an H and let x H. Then x M if
and only if
kx yk kxk

y M

(3.1.13)

Proof. () If x M then, y M x and y are orthogonal and Pythagoras theorem


gives
kx yk2 = kxk2 + kyk2 kxk2

(3.1.14)

() Suppose that Eq(3.1.13) holds. y M and C, then y M and


kx yk2 kxk2 .

(3.1.15)

Rewriting this as an inner product we have


(x y, x y) kxk2
y) (x, y) + ||2 kyk2 kxk2
kxk2 (x,
y) + ||2 kyk2 0.
2Re(x,

(2.1.16)

The last line comes from z C z + z = 2Re{z}. Eq(2.1.16) holds for any
C, it also holds for = tz where t > 0 and z C, where |z| = 1, then
z(x, y) = |(x, y)|. Thus
2t|(x, y)| + t2 kyk2 0
1
k(x, y)| tkyk2 .
2

(2.1.17)

Letting t the result shows (x, y) = 0.


Lemma 3.8. Let M be a closed linear subspace of a Hilbert space H and let x H.
There exists y M z M such that x = y + z.
Proof. Take y M and let it be the closest point to x in M , by definition
kx yk kx mk m M.

(3.1.18)

Let z = x y then x = y + z. m M y + m M and so


kzk = kx yk kx (y + m)k
kz mk m M.

(2.1.19)

By Lemma 2.7, z M .
Theorem 3.9. Riesz-Frechet Theorem
Let H be a Hilbert space and let F be a continuous linear functional on H. There
exists a unique y H such that
F (x) = (x, y)
Furthermore kyk = kF k.

x H.

(3.1.20)

Proof. y is unique because


(x, y) = F (x) = (x, y 0 ) x H

(3.1.21)

implies y = y 0 by Lemma 2.6.


A trivial case of F is when F is the zero operator and we take y = 0.
Let M be the kernel of the linear functional F ,
M = KerF = {x H : F (x) = 0},

(3.1.22)

and M is a proper closed subspace of H. From Lemma 2.8, we know that H =


M M and therefore M 6= {0}. Let z M  z 6= 0. We can scale z with a scalar
such that F (z) = 1. Then pick a z M such that for any x H
x = (x F (x)z) + F (x)z

(3.1.23)

Because H is a direct some of M and its orthogonal complement M , the first


term on the right hand expression is an element of M and the second term is an
element of M . Taking an inner product of both sides with z,
(x, z) = (F (x)z, z) = F (x)kzk2
since z M . If we let y =

z
kzk2

x H,

(3.1.24)

we have
(x, y) = F (x)

(3.1.25)

If kxk 1 then by Cauchy-Schwarz,


|F (x)| = |(x, y)| kxkkyk kyk
Let x =

y
,
kyk

(3.1.26)

which is a unit vector, and therefore


kF k |F (x)| =

|F (y)|
|(y, y)|
=
= kyk.
kyk
kyk

(3.1.27)

Therefore kF k = kyk.

3.2

Tensor Product

Tensor products came around for vector spaces due to its inherent need in physics
and engineering. The following description of tensor products is of vector spaces.
Let R be a commutative ring and M and N be R-modules. We assume that
the rings R will have a multiplicative identity and that the modules are to be unital
1 m = m, m M . The product operation M R N , is calle the tensor product.
However, we will begin with the tensor product of vector spaces first.

Let V and W be vector spaces over a field K, and let {ei } and {fj } be a basis
for V and W respectively. Then the tensor product of these two vector spaces is
V K W defined to be the K-vector space with a basis of formal symbols ei fj ;
simply claim that these new symbols are linearly dependent by definition. Then the
tensors
or elements of the K-vector space are written in formal sums with the form
P
i,j cij ei fj with cij K. Furthermore, v V, w W we define v w to be the
element of V K W given by writing v and w in terms of the original bases of both
V and W then expanding as if the is a non commutative product. For example we
could let V = W = R2 Then we would have a 4 dimensional space R2 R R2 with 4
basis vectors: e1 e1 , e1 e2 , e2 e1 , e2 e2 . If we let v = e1 + 2e2 and w = 3e1 + e2 ,
then
v w = (e1 + 2e2 ) (3e1 + e2 ) = 3e1 e1 + e1 e2 6e2 e1 + 2e3 e3 (3.2.1)
If one were to pick another basis of R2 for the tensor product, we realize that v w
has a meaning in the space R2 R R2 independent of the chosen basis. Looking back
at a more generalized version, with modules, but the properties of the product are
to be satisfied in general, and not just on a basis: for R-modules M and N , their
tensor product M R N is an R-modules spanned, by a spanning set, by all symbols
m n, m M, n N . The symbols satisfy these distributive laws:
(m + m0 ) n = m n + m0 n
m (n + n0 ) = m n + m n0
r(m n) = (rm) n = m (rn)
r Rm, m0 M, n, n0 N

(3.2.2a)
(3.2.2b)
(3.2.2c)

In the space of M R N there are usually more elements than products m n.


The general element of M R N , similar to linear algebra, is called a tensor and is an
R-linear combination
r1 (m1 n1 ) + r2 (m2 n2 ) + ... + rk (mk + nk ),
k 1, ri R, mi M, ni N

(3.2.3a)

Since we know from above that ri (mi ni ) = (ri mi ) ni then let ri mi be renamed
just as mi . Therefore the above linear combination can be written as a sum
k
X

mi ni = m1 n1 + ... + mk nk .

(3.2.4)

i=1

The idea of two sums equal in M R N is not trivial in terms of the given description
of a tensor product. However, there is one case: let M and N be free R-modules with

bases {ei } and {fj } respectively. The resulting space M r N is then also a free module
with basis P
{ei fj }. An element is then equal to another when the coefficients of the
finite sum i,j cij ei fj are the same.
In a general case, where both M and N dont have bases, equality in M R N
relies on using a universal mapping property of the tensor product: M R N is the
universal object that turns bilinear maps on M N into linear maps. Before we
continue to define the tensor product rigorously, we will first define a bilinear map.
Definition 20. A function B : M N P , where M, N , and P are R-modules, is
called bilinear when it is linear in each argument with the other one fixed:
B(m1 + m2 , n) = B(m1 , n) + B(m2 , n)
B(rm, n) = rB(m, n)
B(m, n1 + n2 ) + B(m, n1 ) + B(m, n2 )
B(m, rn) = rB(m, n)
r R, m, m1 , m2 M, n, n1 , n2 N

(3.2.5a)
(3.2.5b)
(3.2.5c)
(3.2.5d)

This idea of bilinearity can be extended further to multilinearity.


Now, any bilinear map M N P to an R- module P can be composed with a
linear map P Q to induce a map M N Q that is bilinear.
bilinear

M N

P
linear

bilinear
Q

The construction of the tensor product of M and N will be the solution to a


universal mapping problem: find an R-module T and bilinear map b : M N T
such that every bilinear map on M N is the composite of the bilinear map b and a
unique linear map out of T . By definition then,
Definition 21. The tensor product M R N is an R-module equipped with a bilinear

B
map M N
M R N such that for any bilinear map M N
P there is a linear
L
map M R N
P making the following diagram commute:

M N
B

M R N
L

Despite the definition M R N is not explicitly given but instead gives a universal
mapping property involving it. Note that maps out of M N are bilinear and those
out of M R N are linear; bilinear maps out of M N turn into linear maps out of
M R N . We can show that any of the two constructions are the same.
Let T, T 0 be R-modules, and b : M N T and b0 : M N T 0 satisfy the
universal mapping property of the tensor product of M and N . From the definition,
the universality of b the map b0 factors uniquely through T : a unique linear map
f : T T 0 . Furthermore, the universality of b0 induces a unique factorization in B
through T 0 : a unique linear map f 0 : T 0 T . Combining the two together we have:
T
b
M N

f
0

T0
f0

b
T

We also notice that there is an unique mapping from f 0 f : T T . This is


also because of the universality of the map b to the R-module T . We can let the
composition map of f 0 f = idT be the identity map. Similarly we have f f 0 = idT 0 .
Therefore both T and T 0 are isomorphic R-modules by f and also f b = b0 , which
means f identifies b with b0 . Any two tensor products of M and N can be identified
with each other in a unique way compatible with the distinguished bilinear maps to
them from M N .
Theorem 3.10. The tensor product of M and N exists.
Proof. The tensor product of M and N , M R N will be the quotient module of a
free R-module. For any set S, we will write the free R-module on S as FR (S) =
L
sS Rs , where s is the term with 1 in the s-position and 0 everywhere else. ie. if
s 6= s0 S rs r0 s0 has r in the s-position, r0 in the s0 -position, 0 elsewhere.
Let S = M N :
M
FR (M N ) =
R(m,n) .
(3.2.6)
(m,n)M N

Let D be a submodule of FR (M N ) spanned by all the elements of:


(m+m0 ,n) (m,n) (m0 ,n) , (m,n+n0 ) (m,n) (m,n0 ) , (rm,n) (m,rn)
r(m,n) (rm,n) , r(m,n) (m,rn) .
(3.2.7a)

The tensor product is then the quotient module by D:


M R N := FR (M N )/D.

(3.2.8)

We write the coset (m,n) +D in M R N as mn. Remembering that the elements


of D are equivalent to 0 in FR (M N )/D gives us
(m+m0 ,n) (m,n) + (m0 ,n)
(m + m ) n = m n + m0 n M R N.
0

(3.2.9a)

The other relations of a bilinear function also hold. This shows that the function
: M N M R N is bilinear. Now, we need to show that all bilinear maps out of
M N factor uniquely through the bilinear map . Suppose P is an R-module and
B : M N P is a bilinear map. Treat M N as a set, so B is just a function on
this set, the universal mapping property of free modules extends B from a function
M N toP to a linear function ` : FR (M N ) P with `((m,n) = B(m, n). We
want to show that ` make sense as a function on M R N ; show that D ker(`)
From the linearity of B,
B(m + m0 , n) = B(m, n) + B(m, n0 ), B(m, n + n0 ) = B(m, n) + B(m, n0 )
rB(m, n) = B(rm, n) = B(m, rbn)
(3.2.10a)
therefore
`((m+m0 ,n) ) = `((m,n) ) + `((m0 ,n) ), `((m,n+n0 ) ) = `((m,n) ) + `((m,n0 ) )
r`((m,n) ) = `((rm,n) ) = `((m,rn) ).
(3.2.11a)
Since ` is linear, these conditions must hold as well:
`((m+m0 ,n) ) = `((m,n) + (m0 ,n) ), `((m,n+n0 ) ) = `((m,n) + (m,n0 ) ),
`(r(m,n) ) = `((rm,n) ) = `((m,rn) ).
(3.2.12a)
Therefore the kernel of ` contains all the generators of the submodule D, so `
will induce a linear map L : FR (M N )/D P where L((m,n) + D) = `((m,n) ) =
B(m, n). Since FR (M N )/D = M R N and (m,n) + D = m n there for the
diagram below commutes as such:

M N
B

M R N
L

This diagram shows that every linear map B out of M N comes from a linear
map L out of M R N such that L(m n) = B(m, n), m M, n N . Now we
show that the linear map L is the only one that makes the diagram commute.
The definition of M R N as a quotient of the free module FR (M N ) tells us
that every element of FR (M N ) is a finite sum
r1 (m1 ,n1 ) + r2 (m1 ,n2 ) + + rk2 (mk ,nk ) .

(3.2.13)

The reduction map FR (M N ) FR (M N )/D = M R N is linear, so every


element of M R N is a finite sum of the form
r1 (m1 n1 ) + r2 (m1 n2 ) + + rk2 (mk nk ).

(3.2.14)

Therefore this means the elements of mn in M R N span it as a R-module. The


linear maps out of M R N are determined by the values of all the elements m n,
thus there is at most one linear map L : M R N P such that m n 7 B(m, n).
Definition 22. Tensors in M R N that have the form m n are called elementary
tensors.
Similar to linear algebra elements of the the tensor product, or tensors, are generally linear combinations of elementary tensors.
We will now prove two simple theorems that run parallel with our idea of linear
algebra.
Theorem 3.11. Le M and N be R-modules with respective spanning sets {xi }iI and
{yj }jJ . The tensor product M R N is spanned linearly by the elementary tensors
xi y j .
Proof. Since an elementary tensor in M R N has the form m n, we write m and
n as spanned by the sets:
X
X
m=
ai x i , n =
bj y j
(3.2.15)
i

where ai = bj = 0 for all but finitely many i and j. Since is bilinear we have
X
X
X
mn=
ai x i
bj y j =
ai b j x i y j
(3.2.16)
i

i,j

which is a linear combination of the tensors xi yj . Since every tensor is a sum


of elementary tensors, the xi yj s span M R N as an R-module.
Theorem 3.12. In M R N , m 0 = 0 n = 0.

Proof. Since m n is additive in n with m fixed, we have


m 0 = m (0 + 0) = m 0 + m 0 m 0 = 0.

(3.2.17)

The same argument is applied to 0 n.

3.3

Tensor products and the Hilbert-Schmidt Class

Now we will look at specific examples of tensor products limited to Hilbert spaces.
In defining the (Hilbert) tensor product H of two Hilbert spaces H1 , H2 , the
approach we take will utilize and emphasize what we did above, the universal
property of the tensor product. In terms of operators, rather than mappings, The
Hilbert space H is characterized, up to isomorphism, by the existence of a bilinear
mapping p : H1 H2 H. It has the following property that each suitable bilinear
mapping L from H1 H2 into a Hilbert space K has a unique factorization L = T p,
with T being a bounded linear operator from H into K.
Before the formal construction of the theory, we need to understand the intuitive
aspects of it. When x1 H1 , x2 H2 , we want to view the element p(x1 , x2 ) H as
a product x1 x2 . The linear combinations of such products form an everywheredense subspace of H. The bilinearity of p implies that these products satisfy certain
linear relations:
(x1 + y1 ) (x2 + y2 ) x1 x2 x1 y2 y1 x2 y1 y2 = 0
x1 , y1 H1 , x2 , y2 H2

(3.3.1a)

All the linear relations satisfied by product vectors can be achieved by use of the
bilinearity of p. The inner product on H satisfies the conditions:
hx1 x2 , y1 y2 i = hx1 , y1 ihx2 , y2 i
kx1 x2 k2 = hx1 x2 , x1 x2 i
= hx1 , x1 ihx2 , x2 i = kx1 k2 kx2 k2
kx1 x2 k = kx1 kkx2 k.

(3.3.2a)

(3.3.2b)

Our construction of the Hilbert space H, the elements of H are complex-valued


functions defined on the product H1 H2 and conjugate-linear in both variables. If
v1 H1 , v2 H2 , v1 v2 is the function that assigns the value hv1 , x1 iv2 , x2 i to the
element (x1 , x2 ) H1 H2 .
Now, suppose that H1 , ..., Hn are Hilbert spaces and is a mapping from the
cartoon product of all these Hilbert spaces into the scalar field C. is called a

bounded multilinear functional on H1 Hn if is linear in each of its variables,


assuming that the other variables remain fixed, and there is a real number c such that
|(x1 , ..., xn )| ckx1 k kxn k,

x1 H1 , ..., xn Hn .

(3.3.3)

If this is so then the lest such constant c is denoted by the norm of the mapping
kk. Then is a continuous mapping from H1 Hn C relative to the product
tor the norm topologies on the Hilbert spaces.
Theorem 3.13. Suppose that H1 , ..., Hn are Hilbert spaces and is a bounded multilinear functional on H1 Hn .
(i) The sum
X
X

|(y1 , ..., yn )|2


(3.3.4)
y1 Y1

yn Yn

has the same finite or infinite value for all orthonormal bases Y1 of H1 ,..., Yn of
Hn .
(ii) If K1 , ..., Kn are Hilbert spaces, Am B(Hm , Km ), (m = 1, ..., n), is a
bounded multilinear functional on K1 Kn , and
(x1 , ..., xn ) = (A1 x1 , ..., An xn )
x1 H1 , ..., xn Hn

(3.3.5)

then
X
y1 Y1

|(y1 , ..., yn )|2 kA1 k2 kAn k2

yn Yn

z1 Z1

|(z1 , ..., zn )|2 , (3.3.6)

zn Zn

when Ym and Zm are orthonormal bases of Hm and Km , respectively where m =


1, ..., n.
Proof. To prove (i) it is sufficient to show that
X
X
X
X

|(y1 , ..., yn )|2

|(z1 , ..., zn )|2 ,


y1 Y1

yn Yn

z1 Z1

(3.3.7)

zn Zn

, whenever Ym , Zm are orthonormal bases of Hm , m = 1, ..., n. Note that since


Ym , Zm are orthonormal bases we can represent the basis vectors in one in terms of a
linear combination of the other.
X
X
y1 =
hy1 , z1 iz1 , ..., yn =
hyn , zn izn
(3.3.8)
z1 Z1

Therefore,

zn Zn

X
y1 Y1

|(y1 , ..., yn )|2 =

yn Yn

y1 Y1

y1 Y1

X
yn Yn

hy1 , z1 iz1 , ...,

z1 Z1

yn Yn z1 Z1

|(

hyn , zn izn )|2

zn Zn

hy1 , z1 i hyn , zn i(z1 , ..., zn )|2

zn Zn

z1 Z1

kz1 k2 kzn k2 |(z1 , ..., zn )|2

zn Zn

z1 Z1

|(z1 , ..., zn )|2

zn Zn

(3.3.9a)
The third line is arrived by noticing that we can rewrite the zm Zm basis in
terms of ym Ym
X
X
z1 =
hz1 , y1 iy1 , ..., zn =
hzn , yn iyn .
(3.3.10)
y1 Y1

yn Yn

Also note that by definition hym , zm i = hzm , ym i and that


kzm k2 = h

hzm , ym iym i

ym Ym

ym Ym

hzm , ym iym ,

hhhym , zm iym , hym , zm iym i

ym Ym

|hym , zm i|2 kym k2

ym Ym

|hym , zm i|2

(3.3.11a)

ym Ym

Using the same argument and exchanging the orthonormal bases we can show
equality.
Then for the proof of (ii), we suppose that 1 m n we choose and fix vectors
y1 Y1 , ..., ym1 Ym1 , zm+1 Zm+1 , ..., zn Zm . The mapping
z (A1 y1 , ..., Am1 ym1 , z, zm+1 , ..., zn ) : Km C

(3.3.12)

is a bounded linear functional on Km , so w Km such that


(A1 y1 , ..., Am1 ym1 , z, zm+1 , ..., zn ) = hz, wi, z Km
From Parsevals equation

(3.3.13)

|(A1 y1 , ..., Am1 ym1 , z, zm+1 , ..., zn )|2

ym YM

|hAm ym , wi|2 =

ym YM

|hym , Am wi|2

ym YM

= kAm wk2 kAm k2 kwk2 = kAm k2

|hzm , wii|2

zm Zm

= kAm k2

|(A1 y1 , ..., Am1 ym1 , z, zm+1 , ..., zn )|2 .

(3.3.14a)

zm Zm

A further summation now yields


X
y1 Y1
2

kAm k

ym Ym zm+1 Zm+1

y1 Y1

|(A1 y1 , ..., Am ym , zm+1 , ..., zn )|2

zn Zn

ym1 Ym1 zm Zm

|(A1 y1 , ..., Am1 ym1 , zm , ..., zn )|2

zn Zn

(3.3.15)
Therefore
X

y1 Y1

|(y1 , ..., yn) |2 =

kAn k2

y1 Y1

X
y1 Y1

y1 Y1

yn Yn

|(A1 y1 , ..., An yn )|2

yn Yn

|(A1 y1 , ..., An1 yn1 , zn )|2

yn1 Yn1 zn Zn

kAn1 k2 kAn k2
X
X
|(A1 y1 , ..., An2 yn2 , zn1 , zn )|2

yn2 Yn2 zn1 Zn1 zn Zn

kA1 k2 kAn k2

X
z1 Z1

|(z1 , ..., zn )|2

(3.3.16)

zn Zn

With H1 , ..., Hn Hilbert spaces, a mapping : H1 Hn C is called a


Hilbert-Schmidt functional on H1 Hn if it si a bounded multilinear functional,
and the sum of Eq(3.3.4) is finite for any choice of the orthonormal bases Y1 in H1 ,
..., Yn in Hn .
Theorem 3.14. If H1 , ..., Hn are Hilbert spaces, the set HSF of all Hilbert-Schmidt
functionals on H1 Hn is itself a Hilbert space when the linear structure, inner
product, and norm are defined by

(a + b)(x1 , ..., xn ) = a(x1 , ..., xn ) + b(x1 , ..., xn )


h, i =

y1 Y1

kk2 =

 X
y1 Y1

(y1 , ..., yn )(y1 , ..., yn )

(3.3.17)
(3.3.18)

yn Yn

|(y1 , ..., yn )|2

 21
,

(3.3.19)

yn Yn

respectively, where Ym is an orthonormal basis in Hm , m = 1, ..., n. The sum


given by Eq(3.3.18) is absolutely convergent, and the inner product and norm do not
depend on the choice of the orthonormal bases Y1 , ..., Yn .
For each v(1) in H1 , ... v(n) in Hn the equation
v(1),...,v(n) (x1 , ..., xn ) = hx1 , v(1)i hxn , v(n)i, (x1 H1 , ..., xn Hn )

(3.3.20)

defines an element v(1),...,v(n) of HSF, and


hv(1),...,v(n) , w(1),...,w(n) i = hw(1), v(1)i hw(n), v(n)i
kv(1),...,v(n) k2 = kv(1)k kv(n)k.

(3.3.21)
(3.3.22)

The set {y(1),...,y(n) : y(1) Y1 , ..., y(n) Yn } is an orthonormal basis of HSF.


There is a unitary transformation U from HSF onto l2 (Y1 Yn ), such that U
is the restriction |Y1 Yn when HSF.
Proof. Choose an orthonormal basis Ym in Hm , m = 1, ..., n, and then associate with
each bounded multilinear functional on H1 Hn the complex-valued function
U obtained by restricting to Y1 Yn . Remember that the condition for a
Hilbert-Schmidt functional is that if is a Hilbert-Schmidt functional if and only if
U l2 (Y1 Yn ).

(3.3.23)

(y1 , ..., yn ) = 0 y1 Y1 , ..., yn Yn .

(3.3.24)

If U = 0, then

Because Ym is an orthonormal basis, its closed linear span is Hm , then it follows


from the multi linearity and continuity of that vanishes throughout H1 Hn .
Let , be Hilbert-Schmidt functionals on H1 Hn , then the same is
true of a + b for all scalars a, b; a + b is a bounded multilinear functional,
U , U l2 (Y1 Yn ) and thus
U (a + b) = aU + bU l2 (Y1 Yn )

(3.3.25)

The sum, given by Eq(3.3.18), can be rewritten in the form


X
(U )(y)(U )(y)

(3.3.26)

yY1 Yn

and it is absolutely convergent with the sum hU , U i, the inner product in


l2 (Y1 Yn ) of the two restricted functionals U and U .
The set HSF of all Hilbert-Schmidt functionals on H1 Hn is a complex vector
space. Eq(3.3.18) then defines an inner product on HSF the restriction U |HSF is a
one-to-one linear mapping from HSF into l2 (Y1 Yn ), and hU , U i = ,
when , HSF. The inner product on l2 (Y1 Yn ) is definite, so is that on
HSF; if HSF and h, i = 0, we have hU , U i = 0, whence U = 0 = 0.
From this we see that HSF is a pre-Hilbert space, and it is apparent from Eq(3.3.18)
that the norm, denoted k k2 in HSF is given by Eq(3.3.19).
From Theorem 3.13, this norm is independent of the choice of the orthonormal
bases Y1 , .., Yn ; this is the same of the inner product on HSF.
Now we will show that U brings HSF onto the whole of the l2 space. Let f
l2 (Y1 Yn ) and xm Hm , m = 1, ..., n, the Cauchy-Schwarz inequality and
Parseval equation gives
X

y1 Y1

|f (y1 , ..., yn )hx1 , y1 i hxn , yn i|

yn Yn

 X

y1 Y1

 21

|f (y1 , ..., yn )|

yn Yn

 X
 21
X
2
2
|hx1 , y1 i| |hxn , yn i|

y1 Y1

= kf k

 X

yn Yn
2

 21

|hx1 , y1 i|

y1 Y1

 X

 21

|hxn , yn i|

yn Yn

= kf kkx1 k kxn k
From this, the equation
X
X
(x1 , ..., xn ) =

f (y1 , ..., yn )hx1 , y1 i hxn , yn i


y1 Y1

(3.3.27a)

(3.3.28)

yn Yn

defines a bounded multilinear functional on H1 Hn , with kk kf k.


The orthonormality of the sets Y1 , ..., Yn , leads to the fact that
(U )(y1 , ..., yn ) = (y1 , ..., yn ) = f (y1 , ..., yn ) y1 Y1 , ..., yn Yn

(3.3.29)

so U = f . Furthermore HSF since U l2 (Y1 Yn ), whence U carries


HSF onto the l2 space.
Since U is a norm preserving linear mapping from HSF onto l2 (Y1 Yn )
completeness of the l2 space entails completeness of HSF; so HSF is a HIlbert space,
and U is a unitary operator.
When v(1) H1 , ..., v(n) Hn , v(1),...,v(n) is a multilinear functional on H1
Hn , and is bounded since
|v(1),...,v(n) (x1 , ..., xn )| kv(1)k kv(n)kkx1 k kxn k

(3.3.30)

by the Cauchy-Schwarz inequality. Furthermore, Parsevals equation gives


X

y1 Y1

X
y1 Y1

 X

yn Yn

|hy1 , v(1)i|2 |hyn , v(n)i|2

yn Yn
2

|hy1 , v(1)i|

y1 Y1

|v(1),...,v(n) (y1 , ..., yn )|2

 X

|hyn , v(n)i|

yn Yn

= kv(1)k2 kv(n)k2 .

(3.3.31)

Hence v(1),...,v(n) HSF and kv(1),...,v(n) k2 = kv(1)k kv(n)k. Using Parsevals


equation again and by absolute convergence,
hv(1),...,v(n) , w(1),...,w(n) i
=

y1 Y1

 X
y1 Y1

v(1),...,v(n) (y1 , ..., y(n))w(1),...,w(n) (y1 , ..., yn )

yn Yn

y1 Y1

hy1 , v(1)i hyn , v(n)ihw(1), y1 i hw(n), yn i

yn Yn


 X

hw(1), y1 ihy1 , v(1)i
hw(n), yn ihyn , v(n)i
yn Yn

= hw(1), v(1)i hw(n), v(n)i

(3.3.32)

When y(1) Y1 , ..., y(n) Yn , the orthonormality of Y1 , ..., Yn implies that


U y(1),...,y(n) is the function that takes the value 1 at (y(1), ..., y(n)) and 0 elsewhere
on Y1 Yn . Therefore
{U y(1),...,y(n) : y(1) Y1 , ..., y(n) Yn }
is an orthonormal basis of l2 (Y1 Yn ), and therefore

(3.3.33)

{y(1),...,y(n) : y(1) Y1 , ..., y(n) Yn }

(3.3.34)

is a basis of HSF.
Let us introduce the notion of the conjugate of a Hilbert space H. For a normal Hilbert space, H has the algebraic structure and inner product defined by the
mappings
(x, y) x + y : H H H
(a, x) ax : C H H
(x, y) hx, yi : H H C.

(3.3.35a)
(3.3.35b)
(3.3.35c)

However we define the conjugate Hilbert space with a slight twist.


is the same set H, with the algeDefinition 23. The conjugate Hilbert space H
braic structure and inner product defined by the mappings:
(x, y) x + y : H H H
(a, x) ax : C H H
(x, y) hx, yi : H H C

(3.3.36a)
(3.3.36b)
(3.3.36c)

where
ax = a
x

and hx, yi = hy, xi.

(3.3.37)

is H.
It is also very easy to see that the conjugate Hilbert space of H
A useful aspect that arises is a subset of a Hilbert space is linearly independent,
orthogonal, or orthonormal, or an orthonormal basis of that space, if and only if it
has the same property relative to the conjugate Hilbert space. If H1 and H2 are
Hilbert spaces and T is a mapping from the set H1 into the set H2 , the linearity of
1 H
2 , and corresponds to conjugateT : H1 H2 is equivalent to linearity of T : H
2 and of T : H
1 H2 . Of course, continuity of T is the same
linearity of T : H1 H
in all four situations, when T is linear the operators have the same bound, since the
j .
norm on Hj is the same as that on H
Definition 24. Suppose that H1 , ..., Hn and K are Hilbert spaces and L is a mapping
from H1 Hn into K. L is a bounded multilinear mapping if it is linear
in each of its variables and there is a real number c such that
kL(x1 , ..., xn )k ckx1 k kxn k.
The least such constant c is denoted by kLk.

(3.3.38)

By a weak Hilbert-Schmidt mapping from H1 Hn into K, we mean a


bounded multilinear mapping L with the properties:
(i)
Lu (x1 , ..., xn ) = hL(x1 , ..., xn ), ui u K

(3.3.39)

is a Hilbert-Schmidt functional on H1 Hn .
(ii) There is a real number d such that kLu k2 dkuk for each u K.
When these conditions are fulfilled, the least possible value of the constant d is
denoted by kLk2 .
A bounded multilinear mapping L : H1 Hn K is (jointly) continuous
relative to the norm topologies on the Hilbert spaces. We see that condition (ii)
actually flows from (i) by an application of the closed graph theorem to the mapping.
Theorem 3.15. Suppose that H1 , ..., Hn are Hilbert spaces.
(i) There is Hilbert space H and weak Hilbert-Schmidt mapping p : H1
Hn H with the following property: given any weak Hilbert-Schmidt mapping L
from H1 Hn into a Hilbert space K, there is a unique bounded linear mapping
T from H into K, such that L = T p; moreover, kT k = kLk2 .
(ii) If H0 and p0 have the properties attributed in (i) to H and p, there is a unitary
transformation U from H onto H0 such that p0 = U p.
(iii) If vm , wm Hm and Ym is an orthonormal basis of Hm , m = 1, ..., n, then
hp(v1 , ..., vn ), p(w1 , ..., wn )i = hv1 , w1 i hvn , wn i,

(3.3.40)

the set {p(y1 , ..., yn ) : y1 Y1 , ..., yn Yn } is an orthonormal basis of H, and


kpk2 = 1.
m be the conjugate Hilbert space of Hm and let H be the set of all
Proof. Let H
1 H
n with the Hilbert space structure given
Hilbert-Schmidt functionals on H
in Theorem 3.14. When v(1) H1 , ..., v(n) Hn , let p(v(1), ..., v(n)) be the Hilbert1 H
n
Schmidt functional v(1),...,v(n) defined on the cartesian product of H
by
v(1),...,v(n) (x1 , ..., xn ) = hx1 , v(1)i hxn , v(n)i
= hv(1), x1 i hv(n), xn i

(3.3.41)

Because we let Yj be an orthonormal basis of Hj , j = 1, ..., n Theorem 3.14 then


says that the set {p(y1 , ..., yn ) : y1 Y1 , ..., yn Yn } is an orthonormal basis of H,
and that
hp(v1 , ..., vn ), p(w1 , ..., wn )i = hw1 , v1 i hwn , vn i
= hv1 , w1 i hvn , wn i,
kp(v1 , ..., vn )k2 = kv1 k kvn k.

(3.3.42a)
(3.3.42b)

From above we have p : H1 Hn H is a bounded multilinear mapping:


we shall prove next that it is a weak Hilbert-Schmidt mapping. Suppose that H,
and consider the bounded multilinear functional p : H1 Hn C defined by
p (x1 , ..., xn ) = hp(x1 , ..., xn ), i.

(3.3.43)

With y(1) Y1 , ..., y(n) Yn , orthonormality of the bases implies that y(1),...,y(n)
takes the value 1 at (y(1), ..., y(n)) and 0 elsewhere on Y1 Yn . Thus
p (y(1), ..., y(n)) = hp(y(1), ..., y(n)), i = hy(1),...,y(n) , i
X
X
=

y(1),...,y(n) (y1 , ..., yn )(y1 , ..., yn )


y1 Y1

yn Yn

y(1)Y1

= (y(1), ..., y(n)),

(3.3.44a)

|p (y(1), ..., y(n))|2 = kk22 .

(3.3.44b)

y(n)Yn

From this we have proven that p is a Hilbert-Schmidt functional on H1 Hn


and that kp k2 = kphik2 ; so p : H1 Hn H is a weak Hilbert-Schmidt mapping
with kpk2 = 1.
Next, suppose that L is a weak Hilbert-Schmidt mapping from H1 Hn into
another Hilbert space K. Let u K and Lu is the Hilbert-Schmidt functional given
in Definition 23, while H and let F be a finite subset of Y1 Yn then we
have using Cauchy-Schwarz
|h

(y1 , ..., yn )L(y1 , ..., yn ), ui|

(y1 ,...,yn )F

|(y1 , ..., yn )||Lu (y1 , ..., yn )|

(y1 ,...,yn )F

 21 

|(y1 , ..., yn )|

(y1 ,...,yn )F

 21

|Lu (y1 , ..., yn )|

(y1 ,...,yn )F


kLu k2

 21

|(y1 , ..., yn )|

(y1 ,...,yn )F


kukkLk2

X
(y1 ,...,yn )F

Hence

|(y1 , ..., yn )|

 21
(3.3.45)

(y1 , ..., yn )L(y1 , ..., yn )k

(y1 ,...,yn )F

kLk2

 21

|(y1 , ..., yn )|

(3.3.46)

(y1 ,...,yn )F

Since
X

y1 Y1

|(y1 , ..., yn )|2 = kk22 <

(3.3.47)

yn Yn

it then follows from Eq(3.3.46) and the Cauchy criterion that the, unordered, sum
X
X

(y1 , ..., yn )L(y1 , ..., yn )


(3.3.48)
y1 Y1

yn Yn

converges to an element T K, and kT k kLk2 kphik2 . Thus T is a bounded


linear operator from H into K, and kT k kLk2 . When y(1) Y1 , ..., y(n) Yn , we
have
T p(y1 , ..., yn ) = T y(1),...,y(n)
=

y1 Y1

y(1),..,y(n) (y1 , ..., yn )L(y1 , ..., yn )

yn Yn

= L(y(1), ..., y(n)).

(3.3.49)

Both L and T p are bounded and multiline and Ym has closed linear span, Hm , m =
1, ..., n then it follows that L = T p.
The condition that T p = L uniquely determines the bounded linear operator T ,
because the range of p contains the orthonormal basis p(Y1 Yn ) of H. u K,
Parsevals equation gives
kLu k22 =

y1 Y1

=
=

X
yn Yn

|hT p(y1 , ..., yn ), ui|2

y1 Y1

yn Yn

y1 Y1

|hL(y1 , ..., yn ), ui|2

|hp(y1 , ..., yn ), T ui|2

yn Yn

= kT uk2 kT k2 kuk2 ;
so we have kLk2 kT k, and thus kLk2 = kT k.

(3.3.50)

We now prove (ii) of the theorem. Suppose that H0 and p0 : H1 Hn H0


have the properties given in (i). When we have that K = H0 and L = p0 , the equation
L = T p0 is satisfied when T is the identity operator on H0 , and also when T is the
projection from H0 onto the closed subspace [p0 (H1 Hn )] generated by the
range p0 (H1 Hn ) of p0 . From the uniqueness of T ,
[p0 (H1 Hn )] = H0

(3.3.51)

kp0 k2 = kLk2 = kT k = kIk = 1.

(3.3.52)

furthermore,

Using a similar argument and letting K = H0 and L = p0 , it follows from the


properties of the Hilbert space H and p given in (i), that theres a hounded linear
operator U : H H such that p0 = U p and
kuk = kLk2 = kp0 k2 = 1

(3.3.53)

The roles of H, p and H0 , p0 can be reversed in this argument, so there is a bounded


linear operator U 0 from H0 into H such that p = U 0 p0 and kU 0 k = 1. Since
U 0 U p(x1 , ..., xn ) = U 0 p0 (x1 , ..., xn ) = p(x1 , ..., xn ),

x1 H1 , ..., xn Hn , (3.3.54)

while
[p(H1 Hn )] = H

(3.3.55)

it follows that U 0 U is the identity operator on H; and similarly U U 0 is the identity


operator on H0 . Finally,
kxk = kU 0 U xk kU xk kxk,

x H

(3.3.56)

so kU xk = kxk, and U is an isomorphism from H onto H0 .

Stinesprings Dilation Theorem

Theorem 4.1. Let A be a unital C - algebra, and let : A B(H) be a completely


positive map. Then there exists a Hilbert space K, a unital -homomorphism : A
B(K), and a bounded operator V : H K with k(1)k = kV k2 such that
(a) = V (a)V

(4.0.57)

Proof. Consider the algebraic tensor product A H, of the unital C -algebra and
the Hilbert Space H, and define a symmetric bilinear function h, i on this space by
setting
ha x, b yi = h(b a)x, yiH ,

(4.0.58)

and extending linearly, where h, iH is the inner product on H. From the definition,
is completely positive thereby ensuring that h, i is positive semidefinite since

X
n
j=1

aj x j ,

n
X


ai x i

i=1


x1
x1 

.. ..

= n ((a a)) . , .
0,
H(n)
xn
xn

(4.0.59)

where h, iH(n) denotes the inner product not the direct sum H(n) of n copies of H,
given by

y1 
 x1
.. ..
= hx1 , y1 iH + + hxn , yn iH
(4.0.60)
. , .
H(n)
yn
xn
A result of positive semidefinite bilinear forms is that they satisfy the CauchySchwarz inequality,
|hu, vi|2 hu, ui hv, vi

(4.0.61)

Therefore we have that


{u A H|hu, ui = 0} = {u A H|hu, vi = 0v A H}

(4.0.62)

is a subspace, N , of A H. The induced bilinear form on the quotient space


A H/N defined by
hu + N , v + N i = hu, vi

(4.0.63)

will be an inner product. We let K denote the Hilbert space that is the completion
of the inner product space A H/N .
If a A, define a linear map (a) : A H A H by
X
 X
(a)
ai x i =
(aai ) xi
(4.0.64)
A matrix factorization shows that the following inequality in Mn (A)+ is satisfied:
(ai a aaj ) ka ak (ai aj ),

(4.0.65)

and consequently,


X


X

(a)
aj xj , (a)
ai x i
X
X
=
h(ai a aaj )xj , xi iH ka ak
h(ai aj )xj , xi iH
i,j

i,j
2

= kak

X

aj x j ,


ai x i .

(0.0.12)

Therefore, (a) leave N invariant and consequently induces a quotient linear transformation on A H/N , which we will still denote by (a). The above inequality also
shows that (a) is bounded with k(a)k kak. Thus, (a) extends to a bounded
linear operator on K, which we will still denote by (a). It is straightforward to verify
that the map : A B(K) is a unital - homomorphism.
Now define V : H K via
V (x) = 1 x + N

(4.0.67)

Then V is bounded, since


kV xk2 = h1 x, 1 xi = h(1)x, xiH kphi(1)k kxk2 .

(4.0.68)

Indeed, it is clear that kV k2 = sup{h(1)x, xiH : kxk 1} = k(1)k.


To complete the proof, we only need to observe that
hV (a)V x, yiH = h(a)1 x, 1 yiK = h(a)x, yiH ,
and so V (a)V = (a).

x, y H,

(4.0.69)

You might also like