Sie sind auf Seite 1von 121

Von Neumann Algebras.

Vaughan F.R. Jones

13th May 2003

in part by NSF Grant DMS9322675, the Marsden fund UOA520, and the Swiss National Science Foundation.

1 Supported

Chapter 1 Introduction.

Chapter 2 Background and Prerequisites


2.1 Hilbert Space
A Hilbert Space is a complex vector space

which is linear in the second variable, satises

H with inner product , : HxH , = , , is positive

denite, i.e.

, > 0
.

for

= 0,

and is complete for the norm dened by test

|| || =

Exercise 2.1.1

Prove the parallelogram identity :

|| ||2 + || + ||2 = 2(|| ||2 + || ||2 )


and the Cauchy-Schwartz inequality:

| , | || || || ||.

Theorem 2.1.2

If C is a closed convex subset of H and is any vector in H, there is a unique C which minimizes the distance from to C , i.e. || || || || C .
Proof. This is basically a result in plane geometry. Uniqueness is clearif two vectors

and

in

minimized the distance,

and these two vectors dene a plane and any vector on the line segment

. d be the distance from C to and choose a sequence n C with ||n || < d + 1/2n . For each n, the vectors , n and n+1 dene a plane. Geometrically it is clear that, if n and n+1 were not close, some point on the line segment between them would be closer than d to .
between and would be strictly closer to To prove existence, let Formally, use the parallelogram identity:

||

n + n+1 2 n n+1 2 || = || + || 2 2 2
5

= 2(||

n+1 2 n 2 || + || || 1/8||n n+1 ||2 ) 2 2 (d + 1/2n )2 1/4||n n+1 ||2 K


such that

Thus there is a constant 2 would be less than d . Thus

||n n+1 ||2 < K/2n C


and has distance

or

n+1 2 || n +2 ||

(n )

is Cauchy, its limit is in

from

Exercise 2.1.3
of

If

(the Banach-space dual of

consisting of all

continuous linear functionals from

H.

Show how to choose a vector

and that

with ( ) = is a conjugate-linear isomorphism from H onto H. orthogonal to

to

C), ker

is a closed convex subset

ker

We will be especially concerned with s eparable Hilbert Spaces where there is an orthonormal basis, i.e. all the a sequence

{1 , 2 , 3 , ...}

of unit vectors with

i , j = 0 i .

for

i=j

and such that

is the only element of

orthogonal to

Exercise 2.1.4

Show that is

linear span of the

{i }

is

{i } is dense in H.

an orthonormal basis for

then the

a : H K is said to be bounded if there is a ||a || K || || H. The inmum of all such K is called the n orm of a, written ||a||. The set of all bounded operators on H is written B (H). Boundedness of an operator is equivalent to continuity. To every bounded operator a between Hilbert spaces H and K, by exercise 2.1.3 there is another, a , between K and H, called the a djoint of a which is dened by the formula a, = , a .
A linear map (operator) number

with

Exercise 2.1.5

Prove that

||a|| =

sup | || ||1,|| ||1

a, | = ||a || = ||a a||1/2 .

The identity map on H is a bounded operator denoted 1. An operator a B (H) is called self-adjoint if a = a . 2 An operator p B (H) is called a projection if p = p = p .

a B(H) is called positive if a, 0 B(H). We say a b if a b is positive. An operator u B (H) is called an isometry if u u = 1. An operator v B (H) is called a unitary if uu = u u = 1. An operator u B (H) is called a partial isometry if u u is a projection.
An operator The last three denitions extend to bounded linear operators between dierent Hilbert spaces.

Exercise 2.1.6 Exercise 2.1.7 Exercise 2.1.8


just sends the

Show that every

a B (H)

is a linear combination of two

self-adjoint operators. A positive operator is self-adjoint. Find an isometry from one Hilbert space to itself that is not H = 2 (N) is a ne example. There is an

unitary. The unilateral shift on obvious orthonormal basis of

indexed by the natural numbers and the shift

nth.
If

basis element to the

(n + 1)th. H
show that the map

Exercise 2.1.9
K
projection.

is a closed subspace of

PK : H

which assigns to any point in

the nearest point in

is linear and a

Exercise 2.1.10
If

Show that the correspondence

ercise is a bijection between closed subspaces of


is a subset of H, S is by denition Note that S is always a closed subspace.

K PK of the previous exH and projections in B (H).

{ H : , = 0 S }. K = K u . u, PK = 1 PK . u H uH
is also

Exercise 2.1.11 Exercise 2.1.12

If If

K u

is a closed subspace then

and

is a partial isometry then so is

The subspace

is then closed and called the initial domain of closed and called the nal domain of the initial and nal domains.
The commutator

the subspace

u.

Show that a partial isometry is the

composition of the projection onto its initial domain and a unitary between

[a, b] K

of two elements of

B (H)

is the operator

ab ba.

Exercise 2.1.13
In general

If

is a closed subspace and

a = a

then

aK K (aK K and

if f

[a, PK ] = 0.

a K K) [a, PK ] = 0.

2.2 The Spectral Theorem


The spectrum

(a)

of

a B (H)

is

{ C : a 1
and that if

is not invertible}.

Exercise 2.2.1
with either

(Look up proofs if necessary.)

empty closed bounded subset of

||a||

or

||a||

in

C (a).

Show that = a , (a)

(a) is a non [||a||, ||a||]

The spectral theorem takes a bit of getting used to and knowing how to prove it does not necessarily help much. If one cannot see the spectral decomposition of an operator it may be extremely dicult to obtainexcept in a small nite number of dimensions where it is just diagonalisation. But fortunately there is nothing like a course in operator algebras, either C or von Neumann, to help master the use of this theorem which is the heart of linear algebra on Hilbert space. The book by Reed and Simon, Methods of mathematical physics vol. 1, Functional Analysis, contains a treatment of the spectral theorem which is perfect background for this course. establish terminology. from the Borel subsets of th (a) (when a = a is normal i.e. [a, a ] = 0) to projections in The spectral theorem asserts the existence of a projection valued measure a or more generally when We will make no attempt to prove it herejust give a vague statement which will

B (H),

written symbolically

E (),

such that

a=

dE ().

This integral may be interpreted as a limit of sums of operators (necessitating a topology on

B (H)), as a limit of sums of vectors a = dE () or simply in terms of measurable functions , a = d , E () . The projections E (B ) are called the spectral projections of a and their images are called the spectral subspaces of a. Given any bounded Borel complex-valued function f on (a) one may form f (a) by f (a) = f ()dE ().
If

is a sigma-nite measure on X and f L (X, ), 2 2 the operator Mf : L (X, ) L (X, ), Mf g (x) = f (x)g (x), is a bounded esssup operator with ||Mf || = (|f (x)|). If f is real valued then Mf is self x X adjoint. Find (f ) and the projection-valued measure E ().

Exercise 2.2.2

Exercise 2.2.3
measure or

If

dim(H) <

nd the spectrum and projection-valued

(which is a Hermitian matrix).

The example of exercise 2.2.2 is generic in the sense that there is a version and of the spectral theorem which asserts the following. If H is any vector a = a B (H), let K be the closed linear span of the {an : n =

0, 1, 2, 3, ...}, then a denes a self-adjoint operator on K and there is a nite measure on the spectrum (a) such that (K, a) is isomorphic in the obvious 2 sense to (L ( (a), ), multiplication by x). Continuing such an argument by restricting to K one obtains a full spectral theorem.
8

Exercise 2.2.4
as functions of

Show that a self-adjoint operator as above.

is the dierence

a+ a
obtained

of two positive operators called the positive and negative parts of

a,

2.3 Polar decomposition


Exercise 2.3.1
square root Show that every positive operator

has a unique positive

1/2

Given an arbitrary

a B (H)

we dene

|a| = (a a)1/2 . u
such that

Exercise 2.3.2
and that

Show that there is a partial isometry

a = u|a|,

ker(a) .

The nal domain of this

is unique subject to the condition that its domain subspace is u is Im(a) = ker(a ) .

2.4 Tensor product of Hilbert Spaces.


H and K are Hilbert spaces one may form H alg K (in the category of complex vector
If one denes the sesquilinear form their algebraic tensor product spaces). On this vector space

, = ,

,
The Hilbert space tensor

and observes that this form is positive denite. product

H K is then the completion of H alg K. It is easy to see that if a B (H), b B (K), there is a bounded operator a b on H K dened by a b( ) = a b .

Exercise 2.4.1
the map

Let

L2 (X, H, )

be the Hilbert space of measurable square

integrable functions (up to null sets) f : X f L2 (X, ) let f L2 (X, H, ) be dened by

f f

denes a unitary from

H. For each H and f (x) = f (x) . Show that H L2 (X, ) onto L2 (X, H, ).

10

Chapter 3 The denition of a von Neumann algebra and some basic examples.
3.1 Topologies on B(H)
(i) The norm or uniform topology is given by the norm

||a||

dened in the

previous chapter. (ii)The topology on by the

B (H)

of pointwise convergence on

strong operator topology. A basis of neighbourhoods of

H is called the a B (H) is formed

N (a, 1 , 2 , ..., n , ) = {b : ||(b a)i || <

i = 1, , n}

(iii)The weak operator topology is formed by the basic neighbourhoods

N (a, 1 , 2 , ..., n , 1 , 2 , .., n , ) = {b : | (b a)i , i | <

i = 1, , n}

Note that this weak topology is the topology of pointwise convergence on

in the weak topology on

dened in the obvious way by the inner

product. The unit ball of of

is compact in the weak topology and the unit ball

B (H)

is compact in the weak operator topology. These assertions follow

easily from Tychono 's theorem.

Exercise 3.1.1

Show that we have the following ordering of the topologies

(strict in innite dimensions). (weak operator topology) < (strong operator topology) < (norm topology)
11

Note that a weaker topology has less open sets so that is a set is closed in the weak topology it is necessarily closed in the strong and norm topologies. We will now prove the von Neumann density or bicommutant theorem which is the rst result in the subject. We prove it rst in the nite dimensional case where the proof is transparent then make the slight adjustments for the general case.

Theorem 3.1.2
with

M be a self-adjoint dim(H) = n < . Then M = M .


Let

subalgebra of

B (H)

containing

1,

M M . y M then y M . To this end we will amplify the action of M on H to an action on HH dened by x( ) = x . If we n choose an orthonormal basis {vi } of H then H H = i=1 H and in terms of matrices we are considering the n x n matrices over B (H) and embedding M
Proof. It is tautological that So we must show that if in it as matrices constant down the diagonal. Clearly enough the commutant of

on

second commutant consists of matrices having a xed element of the diagonal. Let

H H is the algebra of all n x n matrices with entries in M and the M down

n v be the vector n i=1 vi i=1 H and let V = M v H H. Then M V V and since M = M , PV M (on H H) by exercise 2.1.13. So if y M (on HH), then y commutes with PV and yM v M v . In particular y (1v ) = xv for some x M so that yvi = xvi for all i, and y = x M .

Theorem 3.1.3
containing

(von Neumann) Let

be a self-adjoint subalgebra of

B (H)

1.

Then

M =M

(closure in the strong operator topology).

Proof. Commutants are always closed so So let

M M . N (a, 1 , 2 , ..., n , ) be a strong neighbourhood of a. n n We want to nd an x M in this neighbourhood. So let v i=1 H be i=1 i n and let B (H) act diagonally on i=1 H as in the previous theorem. Then the same observations concerning matrix forms of commutants are true. Also M n commutes with PM v which thus commutes with a (on i=1 H). And since 1 M , av = ai is in the closure of M v so there is an x M with ||xi axii || < for all i. aM
and

Corollary 3.1.4
1)

If

M = M

is a subalgebra of

B (H)

with

1 M,

TFAE:

M =M 2) M is 3) M is

strongly closed. weakly closed. A subalgebra of

Denition 3.1.5

B (H) satisfying the conditions of corollary

3.1.4 is called a von Neumann algebra.


12

(A self-adjoint subalgebra of is called a C -algebra.)

B (H)

which is closed in the norm topology

Example 3.1.6
1.

1) Any nite dimensional *-subalgebra of

B (H)

containing

Example 3.1.7 B(H) itself. Example 3.1.8


consider

A =

Exercise 3.1.9 Let (X, ) be a nite measure space and L (X, ) as a *-subalgebra of B (L2 (X, )) as multiplication

operators as in exercise 2.2.2. Show that

and hence a von Neumann algebra. (Hint: if f L and that x = Mf .)

A = A , i.e. A is maximal abelian x A let f = x(1). Show that

Example 3.1.10

If

S B(H), 1
and

we call

(S S )

the von Neumann algebra

generated by S. It is, by theorem 3.1.3 the weak or strong closure of the *algebra generated by

S.

Most constructions of von Neumann algebras

begin by considering some family of operators with desirable properties and then taking the von Neumann algebra they generate. But is is quite hard, in general, to get much control over the operators added when taking the weak closure and all the desirable properties of the generating algebra may be lost.(For instance any positive self-adjoint operator

a with ||a|| 1 is a weak

limit of projections.) However, if the desirable properties can be expressed in terms of matrix coecients then these properties will be preserved under weak limits since the matrix coecients of form

are just special elements of the

, a .

We shall now treat an example of this kind of property which is

at the heart of the subject and will provide us with a vast supply of interesting von Neumann algebras quite dierent from the rst 3 examples. 2 Let be a discrete group and let () be the Hilbert space of all functions 2 f ( )g ( ). An f : C with |f ( )| < and inner product f, g =


orthonormal basis of

()

is given by the

f=

by

f ( )

in the

sense.

{ } where ( ) = , so that For each dene the unitary operator u

(u f )( ) = f ( 1 ). Note that u u = u and that u ( ) = . Thus u is a unitary group representation called the left regular representation. The u are linearly independent so the algebra they generate is isomorphic to the group algebra C. The von Neumann algebra generated by the u goes under various names, U (), () and L() but we will call it vN () as it is the group von Neumann algebra of .
13

To see that one can control weak limits of linear combinations of the

u ,

consider rst the case

= Z/nZ. 0 0 0 0 0 1 1 0 . . 0 0 0 1 0 . . 0

With basis

u0 , u1 , u2 , , un1 , u1

is

represented by the matrix:

0 0 1 0 . .

. 0 0 1 0 .

. . 0 0 1 0 u1
and

which is a matrix constant along the diagonals. Clearly all powers of element of the algebra they generate will have the matrix form (when

all linear combinations of them have this property also so that an arbitrary

n = 6):

a f e d c b

b a f e d c

c b a f e d

d c b a f e

e d c b a f

f e d c b a Z/nZ were replaced

(Such matrices are known as circulant matrices but to the best of our knowledge this term only applies when the group is cyclic.) If by another nite group the same sort of structure would prevail except that the diagonals would be more complicated, according to the multiplication table of the group.

(, ) matrix entry of 1 a nite linear combination of the u 's will depend only on . As observed
Now let It is still true that the above, the same must be true of weak limits of these linear combinations,

be an innite group.

hence any element of

vN (). vN ()
have matrices (w.r.t. the basis {(, ) : 1 is constant}.

We see that the elements of

which are constant along the diagonals :

Exercise 3.1.11

Check whether it should be

or

or some other

similar expression.....
Using the number

on the diagonal indexed by

we can write, formally

at least, any element of

vN ()

as a sum

c u .

It is not clear in what

sense this sum converges but certainly

c u

must dene a bounded linear

operator. From this we deduce immediately the following:

14

(i) The function

is in

. (Apply

c u

to

id .)

(ii)

c u )(

d u ) =

c d1 )u u
on the right hand side con-

verges since

where the sum dening the coecient of c and d1 are in 2 . Exactly what functions Case 1.

dene elements of

vN ()

is unclear but

an important special case gives some intuition.

= Z.

It is well known that the map c n n cnein denes a unitary V 2 2 1 ik from () to L (T). Moreover V un V (e = V un (k ) = V (k + n) = ein eik so that V un V 1 is the multiplication operator Mein . Standard spec tral theory shows that Mein generates L (T) as a von Neumann algebra, 1 and clearly if Mf L (T), V Mf V = cn n where cn ein is the Fourier series of

f.

We see that, in this case, the functions

dene elements of

vN (Z)

are precisely the Fourier series of

c which L functions.

In case we forget to point it out later on when we are in a better position to prove it, one way to characterise the functions which dene elements on vN () is as all functions which dene bounded operators on 2 (). This is not particularly illuminating but can be useful at math parties. At the other extreme we consider a highly non-commutative group, the free group on Case 2.

n generators, n 2. = Fn .

Just for fun let us compute the centre Z (vN ()) of vN (Fn ), i.e. those c u that commute with all x vN (). By weak limits, for c u to be in Z (vN ()) it is necessary and sucient that it commute with every u . This clearly is the same as saying c 1 = c , , i.e. the function c is constant on conjugacy classes. But in Fn all conjugacy classes except that 2 of the identity are innite. Now recall that c is in !! We conclude that c = 0 = 1, i.e. Z (vN ()) = C1. Note that the only property we used of F to reach this conclusion was that every non-trivial conjugacy class is innite (and in general it is clear that

Z (vN ())

is in the linear span of the

with

in a nite conjugacy

class.) Such groups are called i.c.c. groups and they abound. Other examples

S (the group of P SL(n, Z) and Q Q .


include

nitely supported permutations of an innite set),

Unsolved problem in von Neumann algebras: Is

vN (Fn ) = vN (Fm )

for

n=m

( 2)?
and

Note that it is obvious that the group algebras

CFn

CFm

are not iso-

15

morphic, just consider algebra homomorphisms to to

C-but they will not extend C1


is called a

vN ().
A von Neumann algebra whose centre is

Denition 3.1.12
factor.

Exercise 3.1.13 Exercise 3.1.14


M = 1 B (K2 ) M

Show that Suppose

B (H)

is a factor. Show that

so that

H = K1 K2 and let M = B (K1 ) 1 M and M are factors. H.

This exercise is supposed to explain the origin of the term factor as come from a tensor product factorisation of The factor we have constructed as see this consider the function , or

and

B (H). To tr(a) = a1 , 1
from

vN () is of an entirely dierent nature tr : vN () C dened by

tr( c u ) = c1 . This may is clearly linear, weakly 2 continuous, satises tr (ab) = tr (ba) and tr (x x) = |c | 0 (when x = c u ). It is called a trace on vN (). If = Z it obviously equals 2 f ()d under the isomorphism between vN (Z) and L (T). 0

Exercise 3.1.15

tr(ab) = tr(ba), dim H < show that there is a constant K with tr(x) = Ktrace(x). (ii) There is no non-zero weakly continuous linear map tr : B (H) C satisfying tr (ab) = tr (ba) when dim(H) = . (iii) There is no non-zero linear map tr : B (H) C satisfying tr (ab) = tr(ba) and tr(x x) 0 when dim(H) = . (iv) (harder) There is no non-zero linear map tr : B (H) C satisfying tr(ab) = tr(ba) when dim(H) = .
(i)If is a linear map with
Thus our factors

tr : B (H) C

vN ()

when

is i.c.c.

are innite dimensional but

seem to have more in common with of

B (H)

when

dim H <

than when

dim H = ! They H in this case.

certainly do not come from tensor product factorisations

Let us make a couple of observations about these factors. cannot be constant and non-zero down the diagonal. (Take 1)They contain no non-zero nite rank operators for such an operator x x of necessary.)

a |a|

2)They have the property that tr (a) = 0 a = 0 for a positive element 2 (a = |a| by the spectral theorem and all spectral projections of a, hence ,are in the von Neumann algebra generated by a.) 3)They have the property that uu = 1 u u = 1 (i.e. they contain no

non-unitary isometries). Proof. If u u = 1, uu is a projection so 1 tr(u u) = 0.

1 uu

is too and

tr(1 uu ) =

16

Exercise 3.1.16

Show that in

any eld of characteristic

vN (), ab = 1 ba = 1. 0, ab = 1 ba = 1 in F .

Show that if

is

As far as I know this assertion is still open in non-zero characteristic. The above exercise is a result of Kaplansky. The next observation is a remarkable property of the set of projections. 4) If

= Fn , {tr(p) : p

a projection in

vN ()} = [0, 1]. 0 and 1.


To see

Proof. It is clear that the trace of a projection is between

that one may obtain every real number in this interval, consider the subgroup generated by a single non-zero element. By the coset decomposition of Fn 2 the representation of a on (Fn ) is the direct sum of countably many copies of the regular representation. The bicommutant of argument,

ua

is then, by a matrix

vN (Z)

acting in an amplied way as block diagonal matrices

with constant blocks so we may identify

vN (Z)

with a subalgebra of

vN ().

Under this identication the traces on the two group algebras agree. But as we have already observed, any element f L (0, 2 ) denes an element of

vN (Z)

whose integral is its trace. The characteristic function of an interval

is a projection so by choosing intervals of appropriate lengths we may realise projections of any trace. We used the bicommutant to identify

vN (Z) with a subalgebra of vN ().

It is instructive to point out a problem that would have arisen had we tried 2 to use the weak or strong topologies. A vector in () is a square summable 2 sequence of vectors in (Z) so that a basic strong neighbourhood of a on 2 () 2 would correspond to a neighbourhood of the form {b : i=1 ||(a b)i || < } 2 for a sequence (i ) in (Z) with i=1 ||i ||2 < . Thus strong convergence 2 on (Z) would not imply strong convergence on ()! This leads us naturally to dene two more topologies on

B (H).

Denition 3.1.17
{b :
i=1

||(a b)i ||2 < } for any and any sequence (i ) 2 i=1 ||i || < , is called the ultrastrong topology on B (H).
The topology dened by the basic neighbourhoods

The topology dened by the basic neighbourhoods of a, 2 in (Z) with

{b :
i=1

| (a b)i , i | < } (i ), i
in
2

for any

>0

and any sequences

(Z)

with

||i ||2 + ||i ||2 <


i=1

is called the ultraweak topology on

B (H).
17

Note that these topologies are precisely the topologies inherited on if it is amplied innitely many times as

B (H)

B (H) 1K

with

dim K = .

Exercise 3.1.18
B (H)

Show that the ultrastrong topology and the ultraweak topol-

ogy are stronger than the strong topology and weak topology respectively. Show that the ultrastrong and strong topologies coincide on a bounded subset of as do the weak and ultraweak topologies. Repeat the argument of the von Neumann density theorem

Exercise 3.1.19

(3.1.3) with the ultrastrong topology replacing the strong topology.


Here are some questions that the inquisitive mind might well ask at this stage. All will be answered in succeeding chapters. Question 1) If there is a weakly continuous trace on a factor, is it unique (up to a scalar multiple)? Question 2)If there is a trace on a factor p a projection in

is it true that

{tr(p) :

M } = [0, 1]? B (H)?

Question 3) Is there a trace on any factor not isomorphic to

Question 4) Are all (innite dimensional)factors with traces isomorphic? Question 5) If

is a factor with a trace, is

also one? (Observe that

the commutant of a factor is obviously a factor.) Question 6) Is

vN ()

the von Neumann algebra generated by the right

regular representation? Question 7) If

: M N

is a

-algebra isomorphism between von

Neumann algebras on Hilbert spaces so that (a) = uau for a M ?

and

is there a unitary

u:HK

18

Chapter 4 Elementary properties of von Neumann algebras.


Throughout this chapter space

will be a von Neumann algebra on a Hilbert

H. a = a
is an element of

EP1) If projections.

bounded Borel functions of

are in

M , all the spectral projections and M . Consequently M is generated by

all its

According to one's proof of the spectral theorem, the spectral projections

E ()

of

are constructed as strong limits of polynomials in

a.

Or the prop-

erty that the spectral projections of

are in the bicommutant of

may be

an explicit part of the theorem. Borel functions will be in the bicommutant. EP2) Any element in

is a linear combination of

unitaries in

M.

Proof. We have seen that any

a=

and if a is self-adjoint, u+ u . 2

x is a linear combination of 2 self-adjoints, ||a|| 1, let u = a + i 1 a2 . Then u is unitary and M


so that an alternative

EP3)M is the commutant of the unitary group of representation. This follows from EP2)

denition of a von Neumann algebra is the commutant of a unitary group

Exercise 4.0.20
Show that

Show that multiplication of operators is jointly strongly

continuous on bounded subsets but not on all of

B (H).

: B (H) B(H)

is weakly continuous but not strongly con-

tinuous even on bounded sets.


The following result is well known.

19

{a } is a net of self-adjoint operators with a a for and ||a || K for some K R, then there is a self-adjoint a with a = lim a , convergence being in the strong topology. Furthermore a = lub(a ) for the poset of self-adjoint operators.
If

Theorem 4.0.21

a for the limit can be found by weak compactness of the unit ball. Then a , is increasing with limit a, for all H and a a . So lim a a = 0 in the strong topology. Now multiplication is jointly strongly continuous on bounded sets so slim a = a. Note that we have slipped in the notation slim for a limit in the strong topology (and obviously wlim for the weak topology). If a and (a ) are as in 4.0.21 we say the net (a ) is monotone convergent to a.
Prof. A candidate EP4)

is closed under monotone convergence of self-adjoint operators.

The projections on ties:

B (H)

form an ortholattice with the following proper-

p q pH q H p q = orthogonal
projection onto

pH q H

p = 1 p p q = (p q ) = orthogonal
projection onto

pH + q H .

Exercise 4.0.22

Show that

p q = slim n (pq )n . B (H) is complete (i.e.


generate closed under arbitrary

The lattice of projections in

sups and infs) since the intersection of closed subspaces is closed. EP5) The projections in Proof. If

as a von Neumann algebra, and

they form a complete sublattice of the projection lattice of

B (H).

strong limit is

is a set Wof projections in M then nite subsets of S are a p is a net in M satisfying the conditions of 4.0.21. directed set and F pF Thus the strongWlimit of this net exists and is in M . It is obvious that this

pS

p,

the sup being in

B (H).

Easier proof. element of

For each projection

p M , pH

is invariant under each

. Thus the intersection of these subspaces is also. T EP6) Let A be a *-subalgebra of B (H). Let W be ker(a) and K = a A W . Then K is invariant under A and if we let B = {a|K : a A}, then 1K is in the strong closure of

which is thus a von Neumann algebra.

20

Proof. By the above, if

p and q

are projections

is in the strong closure of the algebra theory, if a = a the range projection Pker (a) is in the strong closure of the algebra generated W by a so we may apply the argument of the proof of EP5) P to conclude that is in the strong closure of A. But this is 1K . aA ker(a) Thus if we were to dene a von Neumann algebra as a weakly or strongly closed subalgebra of

p q = 1 (1 p) (1 q ) generated by p and q . By spectral

B (H),

it would be unital as an abstract algebra but its

identity might not be that of

B (H)

so it would not be equal to its bicommu-

tant. However on the orthogonal complement of all the irrelevant vectors it would be a von Neumann algebra in the usual sense. EP7) If

(M p)

and

M is a von Neumann algebra and p M is a projection, pM p = pM p = M p as algebras of operators on pH. Thus pM p and M p

are von Neumann algebras.

pM p and M p commute with each other on pH. Now x (M p) B (pH) and dene x = xp(= pxp) B (H). Then if y M , yx = yxp = ypxp = (xp)(yp) = xpy = x y , so x M and x = px p. Thus (pM ) = pM p which is thus a von Neumann algebra. If we knew that M p were a von Neumann algebra on pH we would be
Proof. Obviously suppose done but a direct attempt to prove it strongly or weakly closed fails as we would have to try to extend the limit of a net in

M p on pH to be in M . So instead we will show directly that (pM p) M p by a clever extension of its elements to H. By EP2 it suces to take a unitary u (pM p) . Let K H be the closure of M pH and let q be projection onto it. Then K is clearly invariant under M and M so q Z (M ). We rst extend u to K by u
for

xi i = u

xi ui
is an isometry:

xi M

and

i H.

We claim that

||u

xi i ||2 =
i,j

xi ui , xj uj

=
i,j

px j xi pui , uj upx j xi pi , uj
i,j

= ... = ||
This calculation actually shows that isometry of So

xi i ||2

K.

By construction

u q M

and

u is well dened and extends to an u commutes with M on M pH,hence on K. u=u qp. Hence (pM p) = M p.
21

Corollary 4.0.23
over the map isomorphism.
Proof.

M is a factor,pM p is a factor on pH, as is pM . Morex xp from M to M p is a weakly continuous *-algebra


If
Continuity and the as before

Factoriality follows immediately from EP7.

*-algebra homomorphism property are obvious so all we have to show is injectivity. But if the projection

xM onto M pH
If or

and is in

xp = 0 then x = 0 on M pH. But Z (M ), hence equal to 1. So x = 0. a M


and

Corollary 4.0.24
implies either
Proof. Let

a=0

M is a b = 0.

factor and

b M

then

ab = 0

p be the range projection of b and apply the previous corollary.


Show that if or

Exercise 4.0.25
a projection in

is a von Neumann algebra generated by the

self-adjoint, multiplicatively closed subset

S,

then

pSp

generates

pM p

(if

is

). Show further that the result fails if

is not closed

under multiplication.

Exercise 4.0.26

W are nite dimensional subspaces of M and M respectively then the map a b ab denes a linear isomorphism between V W and the space V W spanned by all vw with v V and w W .
Show that if is a factor and and
EP8) If and and

a M and a = u|a| is the polar decomposition of a then u M |a| M . Proof. By the uniqueness of the polar decomposition both |a| u commute with every unitary in M . || ||)
is metrizable on

EP9) None of the topologies (except they all are on the unit ball (when except the norm topology. Proof. bounded.

B (H)

but

is separable) and

is separable for all

First observe that a weakly convergent sequence of operators is This follows from the uniform boundedness principle and 2.1.5

which shows how to get the norm from inner products.

{i , i = 1, } be an orthonormal basis of H and let ei be projection onto Ci . Consider the family {em + men : m, n = 1, }. Let V a basic ultrastrong neighbourhood of 0 dened by and {i : ||i ||2 < } and let | |V be the corresponding seminorm, then writing i i 2 i 2 i = j j j we have i,j |j | < . Now choose m so that i |m | < 2 /4 i 2 2 and n so that /4m2 . Observing that ||en (i )||2 = |in |2 we have i |n | <
Here is the cunning trick. Let

|em + men |V |em |V + m|en |V


22

=
i

||em ||2 + m
i

||en ||2

/2 + /2
so that

en + men V . {em + men : m, n = 1, }


can

On the other hand no subsequence of tend even weakly to

since it would have

to be bounded in norm which would force some xed choosing

m to occur innitely

often in the sequence, preventing weak convergence! So by the freedom in

and

to force

em + men

to be in

V,

there can be no countable

basis of zero for any of the topologies (except of course the norm). If we consider the unit ball, however, we may choose a dense sequence i of d(a, b) = [ i 2i ||(a b)i ||2 ]1/2 which is a metric on the unit ball dening the strong topology. (Similarly for the weak topology.) unit vectors and dene We leave non-separability of

B (H)

in the norm topology as an exercise.

EP10) An abelian von Neumann algebra on a separable Hilbert space is generated by a single self-adjoint operator. Proof. Let

{e0 , e1 , e2 , }

be a sequence of projections that is strongly

dense in the set of all projections in the Abelian von Neumann algebra A. 1 Let a = n=0 3n en . The sum converges in the norm topology so a A. The norm of the self-adjoint operator a1 = n=1 is obviously at most 1/2 so that the spectral projection for the interval way we see that all the

[3/4, 2]

for

is

e0 .

Continuing in this

en s are in {a} . This relegates the study of Abelian von Neumann algebras to the spectral

theorem. One can show that any Abelian von Neumann algebra on a sepa rable Hilbert space is isomorphic to either ({0, 1, , n}) (where n = is allowed) or L ([0, 1], dx) or a direct sum of the two. This is up to abstract algebra isomorphism. To understand the action on a Hilbert space, multiplicity must be taken into account.

23

24

Chapter 5 Finite dimensional von Neumann algebras and type I factors.


5.1 Denition of type I factor.
The rst crucial result about factors (remember a factor is a von Neumann algebra with trivial centre) will be the following ergodic property.

Theorem 5.1.1
there is an

If

is a factor and

and

xM

with

pxq = 0.

Moreover

q are non-zero projections in M x can be chosen to be unitary.

Proof. Suppose that for any unitary u M , puq = 0. Then u puq = 0 and W W u pu q = 0. But clearly uM u pu commutes with all unitaries u M u M
so is the identity. The reason we have called this an ergodic property is because of a pervasive analogy with measure-theoretic dynamical systems (and it will become much more than an analogy). serving the measure A transformation 1 is called ergodic if T (A)

T : (X, ) (X, ) pre A implies (A) = 0 or

(X \ A) = 0

A X . If T is invertible one can then show n that there is, for any pair A X and B X of non-null sets, a power T n 2 N of T such that (T (A) B ) = 0. Or, as operators on L (X, ), AT B = 0 where we identify A and B with the multiplication operators by their charn acteristic functions. The proof is the samethe union of all T (A) is clearly invariant, thus must dier from all of X by a set of measure 0.
for a measurable

Corollary 5.1.2
Proof. Let such that

Let

and

there is a partial isometry

q be non-zero projections u (= 0) in M with uu p

in a factor M . Then and u u q .

u be the pxq = 0.

partial isometry of the polar decomposition of

pxq

for

p and q had been swapped around in the proof, so we interchanged uu and u u in the statement.

25

Denition 5.1.3
pM

If

is a von Neumann algebra, a non-zero projection

is called minimal, or an atom, if Show that

(q p) (q = 0 or M
i

q = p).

Exercise 5.1.4
tor.

is minimal in

pM p = Cp.

Denition 5.1.5

A factor with a minimal projection is called a type I fac-

5.2 Classication of all type I factors


We will classify all type I factors quite easily. which we have already seen. We begin with the model,

B (H) 1 be the constant diagonal matrices on H K. Its commutant 1 B (K) will be our model. It is the algebra of all matrices dening bounded
Let operators with every matrix entry being a scalar multiple of the identity matrix on

H.

A matrix with a single

on the diagonal and zeros elsewhere

is obviously a minimal projection.

Theorem 5.2.1
spaces

If

is a type I factor of a Hilbert space

and

and a unitary

u:LHK

with

L, there are Hilbert uM u = B (H) 1.

{p1 , p2 , ...} be a maximal family of minimal projections in M such that pi pj = 0 for i = j . (We assume for convenience that L is separable. Our rst claim is that i pi = 1 so that L = i pi L. For if 1 i pi were nonzero, by corollary 5.1.2 there would be a u = 0 with uu p1 and u u 1 i pi . By minimality uu is minimal and hence so is u u contradicting maximality of the pi . Now for each i choose a non-zero partial isometry e1i with e1i e1i p1 and e1i e1i pi . By minimality e1i e1i = p1 and e1i e1i = pi . Then M is generated by the e1i 's, for if a M we have a = i,j pi apj the sum converging in the strong topology, and pi apj = e1i e1i ae1j e1j p1 M p1 = Cp1 . Thus there are scalars ij so that a = i,j ij e1i e1j . (The details of the convergence of the sum are unimportantwe just need that a be in the
Proof. Let strong closure of nite sums.) If 2

u:

n is the cardinality (X, p1 L) L by

of

{pi }, uf =

let

X = {1, 2, ..., n} e 1i f (i).

and dene the map

i
Observe that

u is unitary and u e1i u is a matrix on 2 (X, p1 L) with an identity operator in the (1, i) position and zeros elsewhere. The algebra generated by 2 2 these matrices is B ( (X )) 1 on (X ) p1 L and we are done.
26

Remark 5.2.2

The importance of being spatial.

We avoided all kinds of problems in the previous theorem by constructing our isomorphism using a unitary between the underlying Hilbert spaces. In general given von Neumann algebras

and

tively, to construct an isomorphism between

N M

generated by and

and

respec-

it suces to construct

(if possible !!!) a unitary u between their Hilbert spaces so that T is contained in uSu . To try to construct an isomorphism directly on S could be arduous at best.

5.3 Tensor product of von Neumann algebras.


If

is a von Neumann algebra on

and

is a von Neumann algebra on

K we dene M N to be {x y : x M, y N }.

the von Neumann algebra on

HK

generated by when you use \h, make sure

it's enclosed in $ signs

Exercise 5.3.1
M alg N

Show that

M N

contains the algebraic tensor product

as a strongly dense *-subalgebra. Let

Denition 5.3.2
units (s.m.u.)

of size

M be a von Neumann algebra . A system of matrix n in M is a family {eij : i, j = 1, 2, ..., n} (n =

allowed) such that (i) (ii) (iii)

e ij = eji . eij ekl = j,l eil


i eii

= 1.
Show that if

Exercise 5.3.3

Neumann algebra 2

M,

then the

{eij ; i, j = 1, ..., n} is eij generate a type I

an s.m.u.

in a von

factor isomorphic to

B ( ({1, 2, ..., n}))

is isomorphic (unitarily equivalent to in this 2 instance) The von Neumann algebra e11 M e11 B ( ({1, 2, ..., n})).

and that

5.4 Multiplicity and nite dimensional von Neumann algebras.


Theorem 5.2.1 shows that type I factors on Hilbert space are completely classied by two cardinalities

(n1 , n2 )

according to:

n1 = n2 =

rank of a minimal projection in rank of a minimal projection in

M , and M.

27

We see that the isomorphism problem splits into abstract isomorphism (determined by isomorphic to A type In factor is by denition one for which

n2 alone), and spatial isomorphism, i.e. unitary equivalence. n = n2 . It is abstractly B (H) with dim H = n. The integer n1 is often called the

multiplicity of the type I factor.


We will now determine the structure of all nite dimensional von Neumann algebras quite easily. Note that in the following there is no requirement that

be nite dimensional.

be a nite dimensional von Neumann algebra on k the Hilbert space H. Then M is abstractly isomorphic to i=1 Mni (C) for some positive integers k, n1 , n2 , ..., nk . (Mn (C) is the von Neumann algebra of all

Theorem 5.4.1
nn

Let

matrices on

Hilbert spaces Ki and a unitary u M u = i B ( 2 (Xi )) 1.


Proof. The centre

n-dimensional Hilbert space.) Moreover there are u : i 2 (Xi , Ki ) H (with |Xi | = ni ) with

Z (M ) is a nite dimensional abelian von Neumann algebra . If p is a minimal projection in Z (M ), pM p is a factor on pH. The theorem follows immediately from theorem 5.2.1 and the simple facts that Z (M ) = k i=1 pi C where the pi are the minimal projections in Z (M ) (Two distinct minimal projections p and q in Z (M ) satisfy pq = 0), and M = i pi M pi .
The subject of nite dimensional von Neumann algebras is thus rather

simple. It becomes slightly more interesting if one considers subalgebras

M.

Let us deal rst with the factor case of this. Let us point out that the

identity of

is the same as that of

N. Im factors are all uniquely M , by the integer (or )

Theorem 5.4.2
k > 0
such that the subfactor
Proof. Let and

If

is a type In factor, its type Ik factor,

determined, up to conjugation by unitaries in

and

pM p is a type mk = n.

being a minimal projection in

N1 and N2 be type Im subfactors with generating s.m.u.'s {eij } {fij } respectively. If k is the integer (in the statement of the theorem) m for N 1 then 1 = 1 eii and each eii is the sum of k mutually orthogonal minimal projections of M , hence n = mk . The same argument applies to N2 . Build a partial isometry u with uu = e11 and u u = f11 by adding
together partial isometries between maximal families of mutually orthogonal projections less than respectively. Then it is easy to check that i ej 1 uf1j is a unitary with wfkl w = ekl . So wN2 w = N1 .

e11

and

f11

w= M

n Now we can do the general (non-factor) case. If N = i=1 Mki (C) and = m j =1 Mrj (C) and N M as von Neumann algebras, let pj be minimal
28

central projections in is a factor and where

and

qi be those of N .

Then for each

p j qi N

is a subfactor so we may form the

(i, j ), pj qi M qi pj matrix = (ij )


by theorem 5.4.2.

ij

is the integer associated with

pj qi N pj qi M qi pj

Exercise 5.4.3
ei is a minimal Mr j C.
Example. Let the form:

Show that the integer projection in the

ij dened factor qi N , ij =

above is the following: if trace of the matrix

pj ei

M = M5 (C) M3 (C) X 0 0 0 X 0

and

be the subalgebra of matrices of

0 0 z

X 0

0 z
is isomorphic to

where and if

z C and X is a 2 2 matrix. Then N p1 = 1 0, q1 = 1 0, etc., we have = 2 1 1 1 .

M2 (C) C

The matrix

is often represented by a bipartite graph with the number

of edges between

and

being

ij .

The vertices of the graph are labelled by

the size of the corresponding matrix algebras. Thus in the above example the picture would be: This diagram is called the Bratteli diagram for

N M.

Exercise 5.4.4
sion

Generalise the above example to show that there is an inclu-

N M

corresponding to any Bratteli diagram with any set of dimen-

sions for the simple components of

N.

5.5 A digression on index.


N M are type I factors we have seen that there is an integer k (possibly ) such that M is the algebra of k k matrices over N . If k < , M is 2 thus a free left N-module of rank k . It seems reasonable to call the number k 2 the index of N in M and write it [M : N ]. This is because, if H < G are groups and CH CG their group algebras, the coset decomposition of G shows that CG is a free left CH -module of rank [G : H ].
If

29

30

Chapter 6 Kaplansky Density Theorem.


6.1 Some simple but telling results on linear functionals.
We begin with a result about linear functionals of independent interest.

Theorem 6.1.1

Let

be a subspace of

B (H)
and

and let

:V C
with

be a linear

functional. T.F.A.E. (i) There are vectors in

H, 1 , 2 , ..., n
n

1 , 2 , ..., n

(x) =
i=1

xi , i

(ii) (iii)

is weakly continuous. is strongly continuous.

(ii) (iii) are obvious, so suppose is strongly continuous. Then let 1 , ..., n be vectors such that ||xi || < 1 i |(x)| < 1 for 2 x V . Then there is a constant K > 0 so that |(x)| < K i ||xi || . Let on V 1( ) by = 1 ...n i H and let K = (V 1)( ). Then dene (i xi ) = (x). Observe that is well-dened and continuous so extends to (x 1)( ) = (x 1)( ), . K and there is a vector = i K with (x) =
Proof. (i)

Exercise 6.1.2

the nite sequences of vectors by

Replace weak and strong by ultraweak and ultrastrong, and 2 -convergent ones in the previous theorem. If

Corollary 6.1.3
sures coincide.

is a convex subset of

B (H),

its weak and strong clo-

31

Proof.

Two locally convex vector spaces with the same continuous linear .

functionals have the same closed convex sets. This is a consequence of the Hahn-Banach theorem to be found in any text on functional analysis.

Corollary 6.1.4

If

dim H =

the strong and ultrastrong topologies dier

the proof here actually on B (H). proves that the weak and ultraweak topologies dier, and to get the result here Proof. Let (i ) be an orthonormal basis of H and let (x) we need to use the previous exercise

1 i n2 xi , i . Then is ultraweakly continuous but not strongly continuous. For if it were n weakly continuous it would be of the form i=1 xi , i and (p) = 0 where p

is the projection onto the orthogonal complement of the vector space spanned by the

i .

But by positivity

(p) = 0

forces

p(i ) = 0

for all

i.

6.2 The theorem


In our discussion of

vN () we already met the desirability of having a normThis is not guaranteed by the von Neumann

bounded sequence of operators converging to an element in the weak closure


of a *-algebra of operators. density theorem. The Kaplansky density theorem lls this gap.

Theorem 6.2.1
the unit ball of

Let

be a *-subalgebra of

B (H).

Then the unit ball of

is strongly dense in the unit ball of the weak closure adjoint part of the unit ball of

of

A,

and the self-

is strongly dense in the self-adjoint part of

M. 1M
and the worried reader may check

Proof. By EP6) we may assume

that we never in fact suppose 1 A. We may further suppose that A is norm-closed, i.e. a C -algebra. Consider the closure of Asa , the self-adjoint The * operation is weakly continuous so if x is a net converging x converges to x so the weak closure x M , x + 2 of Asa is equal to Msa . Let us now prove the second assertion of the theorem. Let x = x M , part of to the self-adjoint element

A.

1 , ..., n , > 0 dene a strong neighbourhood of x. We must 2t come up with a y Asa , ||y || < 1, with ||(x y )i || < . The function t 1+t2 is a homeomorphism of [1, 1] onto itself. So by the spectral theorem we may 2X choose an X Msa with ||X || 1, so that = x. Now by strong density 1+X 2 choose Y Asa with
and

||x|| < 1,

||Y xi Xxi || < ,


Put

and

||

Y X i i || < /4. 2 1+X 1 + X2

y=

2Y and note that 1+Y 2

||y || 1.
32

Now consider the following equalities:

yx= = 2(

2Y 2X 2 1+Y 1 + X2

1 1 2 2 ( Y (1 + X ) (1 + Y ) X ) ) 1+Y2 1 + X2 1 1 Y X = 2( (Y X ) + (X Y ) ) 2 2 2 1+Y 1+X 1+Y 1 + X2 1 1 2 (Y X ) + y (X Y )x. = 2 2 1+Y 1+X 2


By the choice of

Y,

we see that

||(y x)i || <


with

. This proves density for The trick is to form

the self-adjoint part of the unit ball. Now consider a general

x M

||x|| 1.

a b c d is equivalent to strong convergence of the matrix entries so A M2 (C) is a b 0 x strongly dense in M M2 (C). Moreover if strongly c d x 0 a b then b tends strongly to x. And ||b || 1 follows from || || 1 c d a b 0 and b , = , . c d 0 M M2 (C). Strong convergence of a net
to

0 x

x 0

a c

b d

Corollary 6.2.2

If

is a *-subalgebra of

B (H)

containing

then

is a

von Neumann algebra i the unit ball of


Proof. The unit ball of

is weakly compact.

is weakly closed since the norm topology is stronger

than the weak topology, and a closed subset of a compact set is compact. Conversely, if the unit ball of closed. Let net

this proof is wrong!

is weakly compact, then it is weakly there is a

be in the weak closure of

weakly converging to

with

M . By Kaplansky density ||x || 1. Hence x M .

33

34

Chapter 7 Comparison of Projections and Type II1 Factors.


7.1 Order on projections
Denition 7.1.1
isometry If we say that p q u u q . We say that are projections in a von Neumann algebra M if there is a partial isometry u M with uu = p and

and

uM

with

p and q are equivalent, p q uu = p and u u = q .

if there is a partial

Observe that

is an equivalence relation.

Theorem 7.1.2

The relation

is a partial order on the equivalence classes

of projections in a von Neumann algebra .


Proof. Transitivity follows by composing partial isometries. The issue is to show that Let picture: We dene the two decreasing sequences of projections e0 = e, en+1 = and f0 = f, fn+1 = u en u. The decreasing property follows by induc tion since p v pv gives an order preserving map from projections in M

e
and

f v

and

imply

e f.

Compare this situation with sets and

and their cardinalities.

satisfy

uu = e, u u f

vv = f, v v e.

Note the

v fn v

less than

to- .

projections in and

less than

f, v

and

u.

Let

e =
i=0

and similarly interchanging the roles of e ei and f = fi . Note that v f v = e and

f vv f = f so that e f . Also e = (e e1 ) + (e1 e2 ) + + e and f = (f f0 )+(f1 f2 )+ + f are sums of mutually orthogonal projections.
35

i=0

But for each even i, u (ei ei+1 )u = fi+1 fi+2 so ei ei+1 fi+1 fi+2 , and v (fi fi+1 )v = ei+1 ei+2 so one may add up, in the strong topology,
all the relevant partial isometries to obtain an equivalence between

and

f.

Note that if we had been dealing with

vN ()

this argument would have

been unnecessary as we could have used the trace:

tr(v v ) tr(e) = tr(uu ) = tr(u u) tr(f ) = tr(vv ) = tr(v v )


so that

tr(e v v ) = 0

which implies

e = vv.

However in general it is

certainly possible to get a projection equivalent to a proper subprojection of 2 itself. Just take the unilateral shift on B ( (N)) which exhibits an equivalence between

and the projection onto the orthogonal complement of the rst

basis vector. This is analogous to the notion of an innite setone which is in bijection with a proper subset of itself.

Denition 7.1.3
nite if

A projection

pq

for some

p in a von Neumann q < p, p = q . Otherwise p 0.

algebra

is called in-

is called nite. A von

Neumann algebra is called nite if its identity is nite, and it is called purely
innite if it has no nite projections other than

A factor is called innite

if its identity is innite.

We will show that purely innite von Neumann algebras exist though it will not be easy.

Remark 7.1.4 Remark 7.1.5 Remark 7.1.6


For if

If

dim H =

then

B (H)

is innite.

A factor with a trace like

vN ()

is nite.

Every projection in a nite von Neumann algebra is nite.

Or, more strongly, if

pq

and

is nite then

is nite.

pp,
If

p < p,

p=p

then

p + (q p) p + (q p) = q .

Remark 7.1.7
in

M B (H)

if

M is any von Neumann algebra , 1 is an innite projection dim H = .


If

Theorem 7.1.8
or

is a factor and

p, q

are projections in

M,

either

p.
36

Proof.

Consider the family of partial isometries u with uu p, u u q . This set is partially ordered by u v if u u v v and v = u on the initial domain u uH of u. This partially ordered set satises the requirements for Zorn's lemma so let u be a maximal element in it. If u u = q or uu = p we are done so suppose q u u and p uu are both non-zero. Then by 5.1.1 there is a v = 0 with v v q u u and vv p uu . But then u + v is
larger than

which was supposed maximal.

Exercise 7.1.9
tor

Show that two equivalent projections

are unitarily equivalent, i.e. there is a unitary

p and q in a nite facu M with upu = q .

We see that the equivalence classes of projections in a factor form a totally ordered set. are: 1) 2) 3) 4) It is known that, on a separable Hilbert space, the possible isomorphism types for this set

{0, 1, 2, ..., n}

where

n= [0, 1] [0, ] {0, }

is allowed.

type In  type II1  type II  type III

We added 0 to the list for type In , so the comment below about type III and type I1 makes sense

Strictly speaking this is nonsense as type III is the same as type I1 and II1 is the same as II . We mean not only the order type but whether innite or not. Observe that the type II1 case certainly exists. has projections of any trace between We saw that

is

vN (F2 )

and

1.

By the previous theorem

it is clear that the trace gives an isomorphism between the ordered set of equivalence classes of projections ant the unit interval. We will proceed to prove a statement generalising this considerably.

Denition 7.1.10
H
(i) (ii) (iii)

A type II1 factor is an innite dimensional factor

on

admitting a non-zero linear function

tr : M C

satisfying

tr(ab) = tr(ba) tr(a a) 0 tr


is ultraweakly continuous.

The trace is said to be normalised if

tr(1) = 1.

In general a linear functional on a *-algebra A is called positive if (a a) 0 (and (a ) = (a) though this is redundant if A is a C -algebra), and faithful if (a a) = 0 a = 0. A positive is called a

state if

Denition 7.1.11

trace)

1 A and (1) = 1. if (ab) = (ba).

A linear functional

is called tracial (or a

37

It is our job now to show that a II1 factor has a unique ultraweakly continuous tracial state, which is faithful. First a preliminary result on ideals.

Theorem 7.1.12
mann algebra

Let

be an ultraweakly closed left ideal in a von Neu-

M = M e.
Proof.

If

M. M is

Then there is a unique projection 2-sided,

e M

such that

is in

Z (M ).

M M is an ultraweakly closed *-subalgebra so it has a largest projection e. Since e M, M e M. On the other hand if x M let x = u|x| be its polar decomposition. Since u x = |x|, |x| M M . Hence |x|e = |x| and x = u|x| = u|x|e M e. So M = M e. Uniqueness follows easily since f = xe f e. Moreover if M is 2-sided, for any unitary u M , uM = M = uMu = M e = M ueu so ueu = e by uniqueness. Hence e Z (M ).

Corollary 7.1.13
Proof. Let

An ultraweakly continuous positive non-zero trace

Tr

on

a II1 factor is faithful.

M = {x M : T r(x x) = 0}. Then since x a ax ||a||2 x x, M is a left ideal and since T r (ab) = T r (ba), M is a 2-sided ideal. Moreover by the Cauchy Schwarz inequality T r (x x) = 0 i T r (xy ) = 0 y M . Thus M is ultraweakly closed, being the intersection of the kernels of ultraweakly continuous functionals. Thus M = M e for some central projection. And e must be zero since M is a factor.
If

Corollary 7.1.14
projection,

is a type II1 factor on

and

pM

is a non-zero

pM p

is a type II1 factor on

pH . pM p which is non-

Proof. This is cleara trace on mal projection in

restricts to a trace on

zero by faithfulness and all the other properties are immediate. Since a mini-

pM p would be minimal in M , pM p is innite dimensional. tr


will follow easily once we have gathered some facts

The uniqueness of

about projections in a II1 factor.

Theorem 7.1.15
Proof. Let

There are non-zero projections in a type II1 factor of ar-

bitrarily small trace.

d = inf {tr(p) : p M, p2 = p = p = 0}. Suppose d > 0. Let p be a projection with tr(p) d < d. Then p is not minimal since we have seen that M is not isomorphic to B (H). So there is a non-zero projection q < p. But then we have tr(p q ) = tr(p) tr(q ) tr(p) d < d. This is a contradiction. So d = 0.
38

Theorem 7.1.16
positive non-zero
Proof. For Then

M be a type II1 factor with an ultraweakly continuous 2 trace tr . Then {tr (p) : p M, p = p = p} = [0, tr (1)].
Let
consider

S = {p : p a projection in M and tr(p) p is a chain in S , p = p M and p is in the strong closure of the p so p is in S . So by Zorn, S has a maximal element, say q . If tr (q ) were less than r , then by 7.1.8, q p. So choose q = q, q < p. Applying 7.1.14 to p q we nd a projection strictly between q and p. r }. S
is a partially ordered set and if

r [0, tr(1)]

Corollary 7.1.17
interval

The map

tr

gives an isomorphism between the totally or-

dered set of equivalence classes of projections on a type II1 factor and the

[0, tr(1)].

Proof. By 7.1.16 it suces to show that the equivalence class of a projection is determined by its trace. This is immediate from 7.1.8.

Exercise 7.1.18
a subfactor

Let

N M

with

N = Mn (C).

be a type II1 factor . Then for each

nN

there is

Corollary 7.1.19

Any two non-zero ultraweakly continuous normalised traces

on a type II1 factor are equal.


Proof. By the elementary facts it suces to prove that two such traces and

Tr

tr

agree on projections. We may assume one of them, say

tr,

is positive.

By the previous exercise, 7.1.17, and the uniqueness of the trace on a matrix algebra,

tr

and

Tr

are equal on projections for which

tr

is rational. Given

a projection for which

tr(p)

is irrational build an increasing sequence

ei

of

subprojections as follows:

ei with tr(ei ) = T r(ei ) and tr(p) (p ei )M (p ei ) is a type II1 factor so tr and T r agree on projections in it whose tr is arbitrarily close to tr (p ei ). So choose in it a projection ei+1 between ei and p, on which tr and T r agree and with 1 . Then tr and T r agree on tr(p) tr(ei+1 ) < i+1 i ei which is equal to p by the faithfulness of tr .
Suppose we have already constructed Then

tr(ei ) < 1/i.

We shall see that a positive trace on a type II1 factor is norm-continuous and a self-adjoint operator is actually a norm-limit of linear combinations of its spectral projections so in fact an apparently weaker property than ultraweak continuity is all we used in the previous corollarynamely that the trace of the supremum of an increasing net of projections is the supremum of the traces.

39

Corollary 7.1.20
i

Let

be a von Neumann algebra with a positive ultra-

weakly continuous faithful normalised trace

tr.

Then

is a type II1 factor

T r = tr

for all ultraweakly continuous normalised traces

T r.

Proof.

We just have to show that

Z (M )

is trivial.

But if it were not,

choose by faithfulness a projection p Z (M ) with 0 < tr (p) < 1. Dene )tr(xp). Then T r is an ultraweakly continuous normalized T r(x) = ( tr1 (p) trace dierent from tr on 1 p.

be a non-zero positive self adjoint operator. Show + + that there is a bounded piecewise smooth function f : R R such that

Exercise 7.1.21
af (a)

Let

is a non-zero projection. A type II1 factor is algebraically simple. (Hintuse the

Exercise 7.1.22

previous exercise to show that a 2-sided ideal contains a projection, then add projections to obtain the identity.)

7.2 The GNS construction


Thus uniqueness of the trace implies factoriality. This suggests another interesting way to construct a type II1 factor . If in

A = M2 (C), A

is embedded

a a 1. Iterate this procedure to form an An of *-algebras with A1 = A and An+1 = An A, and consider the *-algebra A = n An which could also be called alg,n=1 An . If we normalise the matrix trace on all matrix algebras so that tr (1) = 1 then tr(a 1) = tr(a) so that tr denes a positive faithful normalised trace on A . Elements of A can be thought of as linear combinations of tensors of the form a1 a2 a3 1 1 1 , on which the trace is just the product of the traces of the ai s. We now turn A into a von Neumann
as diagonal matrices: increasing sequence algebra .

AA

x, y = tr(y x). Then A is a preHilbert space and let H be its completion. Note that Mn (C) is a von Neu 2 mann algebra so tr (y x xy ) ||x|| tr (y y ). This means that the operator Lx on A , Lx (y ) = xy , satises ||Lx ( )|| ||x|||| || (where ||x|| is the operator norm of the matrix x and || || is the Hilbert space norm of ) and so extends uniquely to a bounded operator also written Lx on H. One checks that (Lx ) = Lx so x Lx denes a faithful (=injective) representation of the *-algebra A on H . Let M be the von Neumann algebra on H generated by the Lx and identify A with a subalgebra of M .
Dene an inner product on

by

The trace on

is dened by

1 A

considered as a vector in

tr(a) = a, H. So tr extends
40

where

is the element

to a trace on

which is

ultraweakly continuous, positive and normalised. It is also unique with these properties by the uniqueness of the trace on the ultraweakly dense subalgebra

of

M.

If we can show that

tr

is faithful on

then it follows that

is a

type II1 factor . It is important to note that this does not follow simply from the faithfulness of prove it.

tr

on

A.

In fact it is true but we need to do something to

Lx was bounded, the same calculation, with tr(ab) = tr(ba), would have shown that Rx , right multiplication by x, is also bounded. Associativity shows that Lx and Ry commute on A , hence on H. Thus M commutes with Ry for each y A . Now we can show faithfulness: if tr(x x) = 0 for x M then for each a A we have
When we showed that

||x(a)||2 = ||xRa ( )||2 = ||Ra x( )||2 ||Ra ||2 ||x ||2 = ||Ra ||2 tr(x x) = 0.
Since

is dense, this means

x = 0.

So

tr

is faithful on

which is thus a

type II1 factor . Many points are raised by this example. the properties of the vector The easiest to deal with are

which played a prominent role. We used both

M = H

and

M = H.
Let

M be a von Neumann algebra on H. A vector H is called cyclic for M if M = H and separating for M if (x = 0) (x = 0) for all x M .

Denition 7.2.1

Proposition 7.2.2
rating for

With notation as above,

is cyclic for

i

is sepa-

Proof.() Exercisein fact done in the discussion of () Let p be the projection (1 p) = 0 so p = 1. The construction of onto the closure of

above.

M .

Then

pM

. But

from

is a special case of what is known

as the GNS construction (Gelfand-Naimark-Segal). Given a positive lin ear functional satisfying (a ) = (a) on a *-algebra A we let N be {x A : (x x) = 0}. We also dene a sesquilinear form , on A by x, y = (y x). This form is positive semidenite but this is enough for the Cauchy-Schwartz inequality to hold so that

0 y A}

so that

is a subspace and

N is the same as {x : x, y = , denes a pre-Hilbert space


Favourable

structure on the quotient tiplication by

A/N .

Under favourable circumstances, left mul-

denes a bounded linear operator on it. circumstances are provided by C -algebras.

x , Lx

41

is a linear functional on a C -algebra satisfying (a a) 0 show that (a ) = (a). Moreover if A is unital show that is normcontinuous and in fact |||| = (1).
If

Exercise 7.2.3 Remark 7.2.4

It is a standard elementary fact in always adjoin an identity to a C -algebra. If

C -algebras that :AC

one may

Proposition 7.2.5

is a unital

C -algebra

and

is a positive

linear functional then

(y x xy ) ||x||2 (y y )
Proof. Let (1). ||x||2

(a) = (y ay ).

Then

is positive so by the exercise

(x x)

on a unital (x) on the Hilbert space H of the GNS construction via left multiplication: (x)(y ) = xy . Moreover || (x)|| ||x|| and (x ) = (x) since (x)y, z = (z xy ) = y, (x )z . Note that (x) = (x)1, 1 .
It follows immediately that, given a positive linear functional each

C -algebra,

xA

determines a bounded linear operator

To sum up we have the following:

Denition 7.2.6
on

If

is a

C -algebra

and

is a positive linear functional

A,

the Hilbert space of the GNS construction is written

and the rep-

resentation

by left multiplication is called the GNS representation. If

Proposition 7.2.7

is a

C -algebra

on

and

a, . Then is a positive linear functional u : H A such that u (a)u = a.


Proof. Obvious.

and

H, dene (a) = a a denes a unitary

on

A is actually a von Neumann algebra , (A) will not in general be one H . However we will see that if is ultraweakly continuous then (A)
If

will be a von Neumann algebra . For this we need a little preparation.

A be a C -algebra on H containing 1. If is a positive linear functional on A and H is a vector with (i.e. is positive), then there is a t A with = t .
Let
Proof. Dene a sesquilinear form

Lemma 7.2.8

(, ) on A by (a, b ) = (b a). CauchySchwarz and give that |(a, b )| ||a ||||b || so (, ) is well-dened and there is a bounded positive operator t on A with a, tb = (b a). But a, tbc = (c b a) = b a, tc = a, btc so that t A on A . If p = pA , tp is a positive operator in A and if s = t, (a) = a, t = as, s .
.

42

Corollary 7.2.9
(on a

If

on

and

are vectors such that

-algebra

H)

then there is a vector

(a) = a, with = .

is positive

Proof. For

a 0, a, = 1/4( a( + ), + a( ), ) 1/4+ (a).

Theorem 7.2.10

If

is a von Neumann algebra on

and

is a positive

ultraweakly continuous linear functional on M then there are vectors with ||i ||2 < so that (x) = xi , i for x M .
Proof. Let

i H

i , i

be as supplied by 6.1.2. Then for

on

(N), (x) =

x,

so by the previous corollary we are done.

is an ultraweakly continuous positive linear functional on the von Neumann algebra M then the GNS representation is ultraweakly continuous onto a von Neumann algebra on H .
If
Proof. We saw in the last theorem that The map

Corollary 7.2.11

(x) = x 1( ),
By 7.2.7 we

on

x x1

is ultraweakly continuous.

H 2 (N). have that

is ultraweakly continuous since the reduction to continuous. So the kernel of of the form

M 1( )

is ultraweakly

is an ultraweakly closed 2-sided ideal, hence

M e for some e in the centre of M . It follows that is injective on M (1 e) and since the norm of an operator x is determined by the spectrum of x x, the unit ball of the image of M is the image of the unit ball which is
weakly compact so by 6.2.2 we are done.

7.3 Exercises on two projections.


Let

and

be projections onto closed subspaces

and

of the Hilbert

space

respectively. Let

M = {p, q }

Exercise 7.3.1 Exercise 7.3.2

Show that

and this decomposition is

U = (HK)(H K )(HK )(H K)W invariant under p and q .


position, i.e.

W , p and q are in general p q = 0, p q = 1, (1 p) q = 0 and (1 p) q = 1.


Show that, on
43

Exercise 7.3.3

Show that if

a B (H), 0 a 1,

a a(1 a)

is a projection on

H H.

When is it in general position with

a(1 a) 1a 1 0 ? 0 0

a = (p q )2 and A = {a} . Show that a Z (M ) and that {a0 + a1 p + a2 q + a3 pq + a4 qp : ai A} is a *-algebra (which is necessarily weakly dense in M ).
Let

Exercise 7.3.4 Exercise 7.3.5 Exercise 7.3.6

Show that

pM p p

is abelian, generated by

pqp.

>From now on suppose

and

q
in

are in general position.

Show that

p =q

M.

(Hint,consider the polar decomposi-

tion of

pq .)
Show there is a

Exercise 7.3.7
p = e11 .

22

system of matrix units

(eij ) M

with

Exercise 7.3.8

B M2 (C) for some abelian von Neumann algebra B generated by b, 0 b 1, with p correspondb b(1 b) 1 0 ing to and q corresponding to 0 0 b(1 b) 1b
Show that

is spatially isomorphic to

Now drop the hypothesis that

and

are in general position.

Exercise 7.3.9

Show that

pqp =qpq

in

Alternative approach using group representations.

Exercise 7.3.10
group).

Show that

(Z/2Z) (Z/2Z) = Z (Z/2Z)

(innite dihedral

Exercise 7.3.11 Exercise 7.3.12 Exercise 7.3.13

Classify all unitary representations of

Z (Z/2Z).

(Hint

use the spectral theorem for unitaries.) Observe that

2p 1

and

2q 1

are self-adjoint unitaries.

Obtain the structure of 7.3.8 using the last 3 exercises.

44

Chapter 8 The Predual


An ultraweakly continuous linear functional

on a von Neumann algebra M is norm continuous so denes an element of M . Our goal in this chapter is to show that the set of all such is a closed subspace M of M and that the duality between M and M makes M equal to the Banach space dual of M . We will rst establish this in the special case M = B (H).

8.1 Trace class and Hilbert Schmidt operators.


The material in this section is standard so we will only prove results as it suits us, otherwise referring any unproved assertions to Reed and Simon.

Lemma 8.1.1
bases of

If

a B (H) is positive and (i ) and (i ) are two orthonormal ai , i =


i i

H,

then

ai , i

(where

is a possible value for the sum).

Proof. We have

ai , i =
i i

|| ai ||2 ai , j |2 ) aj , i |2 )

=
i

(
j

| |
i

=
j

( =
j

|| aj ||2
45

=
j

aj , j

where every number is positive so the order of the sum is immaterial. The number written

ai , i

of the previous theorem is called the trace of

a,

T race(a).
An element

Denition 8.1.2
.
If

a B (H) is said to be of

trace class if

T race(|a|) <

is trace class and

(i )

is an orthonormal basis, the sum

ai , i
i
converges absolutely and is called the trace,

T race(a), H

of

a. |a|1 = ||a||

Theorem 8.1.3
T race(|a|) |a|1 .

The trace class operators on

form a self-adjoint ideal

of compact operators,

I1 ,

in

denes a norm on

B (H). The I1 for which

function

|a|1

dened by

it is complete. Moreover

Proof. The only thing not proved in Reed and Simon is completeness. For this observe that if

an

is a Cauchy sequence in

| |1 ,

what we have to do is show that the norm limit of a

(an )

is trace class and that the sequence tends to

|| || so | |1 -Cauchy sequence that limit in | |1 . So


it is Cauchy in

suppose

>0

is given. Then for

and

large enough

|an am |i , i < .
i=1
So for any

N,
N

|an am |i , i < .
i=1
Now if bn tends in norm to b, then |bn | tends in norm to |b| (obviously bn bn b b, and approximate the square root function by polynomials on an interval) so for each xed i,

n
So

lim |an am |i = |a am |i . a I1
since

N i=1 |a am |i , i < and letting N tend to we see that I1 is a vector space, and also that an a in | |1 .
46

The trace is independent of the orthonormal basis and if and

a h

is trace class

b B (H), T r(ab) = T r(ba). h I1


determines a linear functional on

We see that each

B (H)

by

h (x) = T race(xh).

Denition 8.1.4
for the state

The trace-class matrix as above is called the density matrix

h .
Each

Proposition 8.1.5
element of

is ultraweakly continuous and its norm as an

B (H)

is

|h|1 .

Proof. Since of

h is compact, choose an orthonormal basis (i ) of eigenvectors |h| with eigenvalues i and let h = u|h| be the polar decomposition. Then

h (x) =
i=1

xu|h|i , i

so ultraweak continuity is apparent, and

h (x)
i=1

||x|| || |h|i ||

= ||x||
i=1

= ||x|| |h|1 .
Moreover evaluating If

on

gives

||h || = |h|1 .

Hilbert-Schmidt if

are Hilbert spaces, a bounded operator x : H K is called 2 x x is trace class, i.e. i=1 ||xi || < for some (hence any) orthonormal basis (i ) of H. The set of all Hilbert-Schmidt operators 2 from H to K is written (H, K) and if x is Hilbert-Schmidt, so is x , and x and is compact.

Theorem 8.1.6

a B(H), b B(K) and x 2 (H, K) then bxa 2 (H, K). If x (H, K) and y 2 (K, H) then yx is trace class. With the 2 inner product x, y = T race(y x), (H, K) is a Hilbert space in which the
If 2
Proof. See Reed and Simon.

nite rank operators are dense.

Exercise 8.1.7

Prove all the assertions made above about trace-class and

Hilbert-Schmidt operators.
47

Exercise 8.1.8
K H
Let

to

H and K are Hilbert spaces construct (H, K) and show that it is unitary.
If

a natural map from

|x|2

be the Hilbert space norm on Hilbert-Schmidt operators.

Lemma 8.1.9
For then let basis

If

(H, K) and y

(K, H) then T race(xy ) = T race(yx). |x| is trace class.

Proof. First note that the result is true if we suppose that

x = u|x|

be the polar decomposition, choose an orthonormal

of the nal domain of u and extend it to an orthonormal basis of K. Also extend (u i ) to an orthonormal basis of H by vectors in ker(|x|). Then

(i )

T race(xy ) =
i

u|x|yi , i

=
i

|x|yuu i , u i = T race(|x|(yu)) = T race((yu)|x|) = T race(yx.)

Now suppose only that

of nite rank with

x is Hilbert-Schmidt. |x x |2 < . Then

Let

>0

be given and choose

|T race(xy ) T race(yx)| = |T race((x x )y ) T race(y (x x ))|


which by Cauchy-Schwartz is

2 |y |2 . = h .

Corollary 8.1.10
B (H),

If

is an ultraweakly continuous linear functional on

there is a trace class

so that

(i ) and (i ) in 2 (N, H) so that (x) = i xi , i . 2 Then if we dene a and b from (N) to H by a(f ) = i f (i)i and b(f ) = i f (i)i , a and b are Hilbert Schmidt and (x) = T race(b xa) which is T race(xab ) by the previous result.
Proof. By 6.1.2 there are Putting everything together so far, we have identied the image of the Banach space I1 under the map h h with the closed subspace of B (H) consisting of ultraweakly continuous linear functionals. To close the loop we only need to show that the Banach space dual of

I1

is

B (H).
for

Theorem 8.1.11
x B (H)
so that

: I1 C is linear and bounded (a) = a (x), and |||| = ||x||.


If
48

| |1 ,

there is an

Proof. This is rather routine. Two vectors

x of I1 by x(v ) = v, so one may dene a sesquilinear form on H by (, ) = (x). Boundedness of x follows from that of so there is an appropriate x B (H). To show that the norm of x as an element of the dual of I1 is actually ||x||, suppose ||x|| = 1 and choose a unit vector with ||x || almost equal to 1. Then T r (hx) is almost 1 if h is the partial isometry which sends v H to . v, x ||x ||
and dene an element

Exercise 8.1.12

Fill in the missing details in the previous proof.

Now we pass to von Neumann algebras though in fact these results work for any ultraweakly closed subspace of

B (H). B (H)
then

Theorem 8.1.13
V

If

is an ultraweakly closed subspace of

in the sense that if

which

(V ) = 0

then

(x) = 0 xV.

for every ultraweakly continuous

V = for

Proof. This is a simple application of the Hahn-Banach theoremif

x /
and

construct an ultraweakly continuous functional which is zero on

non-zero on

x.
Exhibit a non-zero trace class operator on
2

Exercise 8.1.14
orthogonal to

()

which is

vN ().
If

Theorem 8.1.15
topology on

is an ultraweakly closed subspace of

B (H)

then it

is canonically the dual Banach space of

which is dened as the space

of ultraweakly continuous linear functionals on

V.

Moreover the ultraweak

is the weak-* topology on

as the dual of

V .

is a Banach space with dual B and V is a weak-* closed subspace of B then V is the dual of B/V (surjectivity of the natural map from V to the dual of V /B is a result of the previous
Proof. If

theorem), so V is a dual space. So we just have to identify the Banach space B/V with the space of weak-* continuous (as elements of B ) linear functionals on

V.

This is a simple exercise. Putting

B = I1

we are done.

Exercise 8.1.16
V

If

is an ultraweakly closed subspace of

B (H),

show that

is a separable Banach space if

is a separable Hilbert space.

49

8.2 Normal maps


There is still a very unsatisfactory feature about the previous section in that ultraweak continuity and the predual depend on the Hilbert space representation on which a von Neumann algebra acts. In this section we shall show that ultraweak continuity is actually equivalent to a property called normality which is determined by the purely algebraic structure of a von Neumann algebra .

Denition 8.2.1

A positive linear map

:AB

between von Neumann

algebras is called normal if

a ) =

(a ) A.

for any increasing net

(a )

of self-adjoint operators in

Theorem 8.2.2

A positive linear functional on a von Neumann algebra is

normal i it is ultraweakly continuous.

Proof. That ultraweak implies normal is obvious. For the moment we refer to Dixmier for the other direction and content ourselves to remark that our proof of the uniqueness of the trace on a II1 factor used only normality (a self-adjoint operator is a norm limit of linear combinations of its spectral projections) so that a normal trace is attached to it by its algebraic structure alone. Observe that by 6.1.2 any ultraweakly continuous linear functional is a linear combination of positive ones (by polarisation) so the predual of a von Neumann algebra is dened entirely by the algebraic structure. In fact the predual is unique up to isometry. There is another important structure theorem for whose proof we will also provisionally simply refer to Dixmier:

Theorem 8.2.3

be a normal homomorphism of von Neumann algebras

is the composition of an amplication

x x id,

a reduction to a closed

invariant subspace, and a unitary equivalence.

Proof. Dixmier. A unitary equivalence is also called a spatial isomorphism.

50

8.3 A technical lemma.


Let us prove a lemma which shows what the techniques developed so far can be good for. It will be crucial in our treatment of Tomita-Takesaki theory. It is a Radon-Nikodym type theorem inspired by one due to Sakai([]).

Lemma 8.3.1
1/2)
so that

Let

R+

be given and let

be a faithful ultraweakly con-

tinuous state on a von Neumann algebra M . | (y x)| (x x) (y y ). Then there is an

M be such that a M1/2 (elements of norm


Let

(x) = (ax) + 1 (xa).


Proof. For

aM

let

a (x) = (ax + 1 xa). M

Then the map

(a) = a , (M1 ) (M ).

is continuous for the topologies of duality

: M M , between M and M .
is not in

But we know that this topology on

is the ultraweak topology so that

is a compact convex set. By contradiction suppose that

Then by Hahn-Banach there is an of

D = supaM1/2 (a (h)). But if h, we have u /2 (h) = 1/2((|h|) + 1 (|h |))


so that

h M with ( (h)) > D where h = u|h| = |h |u is the polar decomposition

1 2D (|h|) + (|h |) 2 (|h|) (|h |). D < | (h)| = | (u|h|1/2 |h|1/2 )| (|h|) (u|h|u ),
a contradic-

But also tion.

51

52

Chapter 9 Standard form of a II1 factor and II factors.


9.1 Standard form.
In this section

tinuous faithful normalized trace L2 (M ).

will be a von Neumann algebra with an ultraweakly contr and L2 (M, tr) will be abbreviated to

In section 7.2 we learned how to construct a von Neumann algebra from a

-algebra and a positive linear functional on it. If we apply this construction to L (X, ) with trace given by f d, the Hilbert space would be L2 (X, d). 2 For this reason, if M is a type II1 factor we write L (M, tr ) for the GNS p Hilbert space obtained from the trace. In fact one can dene L spaces p 1/p for 1 p using the L norm ||x||p = tr (|x|) . A noncommutative p version of the Holder inequality shows that || ||p is a norm and L (M ) is 1 the completion. We set L (M ) = M and we shall see that L (M ) is the predual of

M . for the vector in L2 (M ) which is the identity || ||-unit

Let us x on the notation

M.
If

Proposition 9.1.1

is a type II1 factor the

ball of

is a

complete metric space for

|| ||2

and the topology dened by

|| ||2

on the

unit ball is the same as the strong (and ultrastrong and *-strong) topology.

xn is Cauchy in || ||2 then for each a M , xn a is also since ||xa||2 ||a|| ||xn ||2 . So we can dene x on the dense subspace M of L2 (M ) by x(a) = limn xan . Since ||x|| 1, we have ||x || || || for M 2 so x extends to a bounded operator on L (M ) which is obviously in M , and x = x = limn = limn xn in || ||2 .
Proof. If 53

The strong topology is obviously no stronger than seminorm

|| ||2

since the single

a ||a||

denes the

|| ||2

topology. Moreover

||xa|| ||x||2 ||a||

shows that

|| ||2

controls

the strong topology on the unit ball. Finally note that in the statement of the theorem it does not matter what representation of

is used to dene the strong topology on the unit ball as

the ultrastrong topology does not change under the manipulations that we used to get the GNS construction from a II1 factor on an arbitrary Hilbert space.

2 The action of M on L (M, tr ) is called the standard form of M . Note 2 that vN () on () is already in standard form. (We see that we could have
obtained our rst example of a II1 factor by applying the GNS construction to the group algebra form.

c u = ci d.) We now want to determine the commutant M when M is in standard


with the trace

tr(

Denition 9.1.2
tion which is the

J : L2 (M ) L2 (M ) be the antilinear unitary involu2 extension to L (M ) of the map x x from M to M .


Let

Lemma 9.1.3
(i) (ii)

For

x, a

in

M,

and

in

L2 (M )

J, J = , JxJ (a ) = ax
Proof.

(i) If (ii)

= a and = b, J, J = tr(ba ) = , JxJ (a) = J (xa ) = ax .


For

Corollary 9.1.4 Lemma 9.1.5


Proof. Take

on

L2 (M ), JM J M

Proof. Left and right multiplication commute.

For

on

L2 (M ),

if

x M , Jx = x .

a M,

then

Jx, a = Ja, x = a , x = , xa = x , a .

54

Theorem 9.1.6

For

on

L2 (M ), JM J = M

Proof. If we can show that x x, is a trace on M we are done since 2 2 by the above results L (M ) can be canonically identied with L (M ) so that the two also for But for

J maps coincide. (Note that is cyclic and M .) So we would have JM J M . x, y M , xy , = y , x = y , Jx = x, Jy = x , y = yx, .

separating for

hence

We see that the commutant of the left regular representation of since

on

()

is the von Neumann algebra generated by the right regular representation

Ju J = 1 . And more generally the commutant of the left action 2 of M on L (M ) is the algebra of right multiplication operators. In 2 particular the commutant of a type II1 factor M on L (M ) is also a type II1 factor . This is not the case for M on an arbitrary Hilbert space. For 2 instance we could consider M 1 on L (M ) H for some innite dimensional H. Then the commutant of M 1 would be JM J B (H)innite matrices over JM J .

Denition 9.1.7
type II1 factor and

A II factor is a factor on the form

M B (H)

with

dim H = .
Let

Proposition 9.1.8
that

be an innite factor with a projection

pM

so

pM p

is a type II1 factor . Then

is a II factor.

{p } of mutually orthogonal projections in M with Put q = p . If 1 p were p then we could contradict the maximality of the family {p }. So write 1 = q + p with q p. Since M is innite the set of indices {} is innite so we may choose a bijection with itself minus 0 and write q + p 0 + =0 p = p 1. We conclude that p is equivalent to 1 so we may suppose it equal to 1. We may then construct a system of matrix units by using partial isometries implementing the equivalences between the p to obtain the result
Proof. Choose a maximal family

p = p .

from exercise 5.3.3.

55

Note that we used implicitly in the above proof the fact that the supremum of nitely many nite projections is nite. This will be covered in a series of exercises on projections. It could conceivably happen that, given a II factor of the form

M , the type II1 factor

pM p

depends on

(obviously only up to equivalence). We now

introduce the trace on a II factor which will make this issue more clear. If

is a type II1 factor , dene the map

tr
by

from

(M B (H))+

(the set

of positive elements of

M B (H)),

to

[0, ]

tr((xij )) =
i=1

tr(xii ) H
to

where we have chosen a basis of the innite dimensional Hilbert space identify

M B (H)

with certain matrices over

M.

M be as above. (i) tr (x) = tr (x) for 0. (ii) tr (x + y ) = tr (x) + tr (y ). (iii) If (a ) is an increasing net of positive operators with a = a then tr( a ) = lim tr(a ). (iv) tr (x x) = tr (xx ) x M . (v) tr (uxu ) = tr (x) for any unitary u M and any x 0 in M . (vi) If p is a projection in M then p is nite i tr (p) < . (vii) If p and q are projections with p nite then p q i tr(p) tr(q ). (viii) pM p is a type II1 factor for any nite projection p.
Let
Proof. The rst two assertions are immediate. For (iii), note that the

Theorem 9.1.9

diagonal entries of positive matrices are ordered as the matrices, and all numbers are positive in the sums. (iv)Is obvious using matrix multiplication. (v) follows from (iv) via uxu = (u x)( xu ). For (vi), if tr (p) < but p is innite, there is a proper subprojection of If

p having the same trace as p.

The

dierence would be a projection of trace zero which is clearly impossible.

tr(p) = then if q is a projection of nite trace, q p and if q p then tr (p q ) = so one may construct an innite sequence of mutually orthogonal equivalent projections less than p. Using a bijection with a proper subsequence, p dominates
an innite projection so is innite itself. (vii) follows easily as in the case of a type II1 factor . For (viii) simply observe that

tr(p) < q.

means that

for some

whose matrix is zero except for nitely many

1's

on the

diagonal. And obviously

qM q

is a type II1 factor for such a 56

Corollary 9.1.10
tr M
to the interval

Let

be a II factor on a separable Hilbert space and

be the trace supplied by a decomposition II1

B (H).

Then

tr

denes an

isomorphism of the totally ordered set of equivalence classes of projections in

[0, ].

Proof. Given the previous theorem, we only have to prove that any innite projection is equivalent to the identity. But if p is innite choose u with uu = p and u u strictly less than p. Then (u )n u are a strictly decreasing sequence of equivalent projections so we may write p as an orthogonal sum p = p + i=1 pi with all the pi equivalent for i 1. Now write the identity as a countable orthogonal sum of projections all II1

p1

(using the decomposition

B (H)

if necessary). We see that

1 p.

B (H) case, the trace cannot be normalised (by tr (1) = 1 in the type II1 factor case or tr (minimalprojection) = 1 in the B (H) case). This allows for the possibility of an automorphism of M with tr ((x)) = tr (x) for x 0 and > 0, = 1.
Unlike the II1 case, or for that matter the

Exercise 9.1.11 Exercise 9.1.12

Show that the trace on a II factor is unique with proper-

ties (i) to (vi), up to a scalar. If

onto another, then

: M N is a *-homomorphism from a type II1 factor is an isomorphism, strongly continuous on the unit ball.

57

58

Chapter 10 The Coupling Constant


We want to compare actions of a given II1 factor on (separable) Hilbert spaces. We will show that they are parameterized by a single number in

[0, ].

Denition 10.0.13
Hilbert space from

If

is a type II1 factor , by

M -module

we will mean a

to a type II1 factor acting on

write that

H together with an ultraweakly continuous unital *-homomorphism H. Thus M acts on H and we will action simply as x for x M and in H.
The identity

In fact the ultraweak continuity condition is superuous. map makes the Hilbert space on which

M is dened into an M -module. Given M on H and another Hilbert space K, x x id makes H K into 2 an M -module. The GNS representation makes L (M ) into an M -module. (The notion of M M bimodule is dened similarly as two commuting actions 2 of M on some Hilbert space, L (M ) being the rst example.) There is an obvious notion of direct sum of M -modules. We will compare a given M 2 module H with L (M ) by forming the direct sum of it H and innitely many 2 copies of L (M ).

10.1 Denition of dimM H


Theorem 10.1.1
(i.e. Let

Then there is an isometry

be a type II1 factor and 2 2

u : H L (M )

H a separable M -module. (N) such that ux = (x 1)u

is

M -linear).

M -module K = L2 (M ) 2 (N). Let p = id 0 B (K) 2 2 be the projection onto H and q = 0 id be the projection onto L (M ) (N). Both p and q are in M (on K) which is a II factor since q is clearly innite in
Proof. Form the 59

e is a rank one projection in B ( 2 (N)) then (0 (1 e))M (0 (1 e)) 2 is a type II1 factor , being the commutant of M on L (M ). Since q is an innite projection in M , by 9.1.10 there is a partial isometry in M with u u = p and uu q . Using the obvious matrix notation for operators on K, let u be represented by
and if

a b c d
Then calculating so that

u u = p and uu q gives b b+d d = 0 and aa +bb = 0 0 w 0 0

u=
for some isometry Moreover the fact

w : H L2 (M ) 2 (N). that u commutes with M

is equivalent to

wx = (x 1)w

x M .

Corollary 10.1.2

The commutant of a type II1 factor is either a type II1

factor or a type II factor.


Proof. We leave the proof as an exercise.

Proposition 10.1.3
then

uu M

on

u : H L2 (M ) 2 (N) is an M -linear L (M ) 2 (N) and tr(uu ) is independent of u.


If
2

isometry

v were another M -linear isometry then uu = uv vu tr(uu ) = tr((vu )(uv )) = tr(vv ).


Proof. If

so by 9.1.9

Observe that if M were replaced by C in the above construction the number tr (uu ) would be the dimension of H.

For a type II1 factor (or the n n matrices) and an M module H, the number tr (u u) dened by the two previous results is called

Denition 10.1.4

dimM H,
whose

or the coupling constant or the

M -dimension

of

H.

Simply by reducing by projections in

M -dimension

is any number in

(M 1) [0, ].

one obtains Hilbert spaces

Trivial examples

dimM L2 (M ) = 1. 2 2 (ii) dimM (L (M ) (N)) =


(i) 60

10.2 Elementary properties of dimM H


Theorem 10.2.1
(i) With notation as above,

dimM (H) < i M is a type II1 factor . (ii) dimM H = dimM K i M on H and M on K
spatially isomorphic). (iii) If

are unitarily equivalent (=

Hi

are (countably many)

M -modules, dimM Hi .
i

dimM (i Hi ) =
(iv) (v)

dimM (L2 (M )q ) = tr(q ) for any projection q M . 1 If p is a projection in M , dimpM p (pH) = trM (p) dimM (H). M
is nite, hence a type II1 factor

For the next two properties we suppose with trace (vi) If (vii)

trM

p is a projection in M , dimM p (pH) = trM (p) dimM H. (dimM H)(dimM H) = 1.


Using an

Proof.

equivalent to

on

M -linear isometry u we see that M on H is unitarily uu L2 (M ) 2 (N). This makes (i) and (ii) obvious. M -linear
isometries

To see (iii), choose

ui

from

Hi

to

L2 (M ) 2 (N)

and

compose them with isometries so that their ranges are all orthogonal. Adding we get an M -linear isometry u with uu = ui u i . Taking the trace we are done. For (iv), choose a unit vector

uu

(N)

and dene

u(v ) = v .
where

Then

is

JqJ e

where

is a rank one projection.

(v) Let us rst prove the relation in the case projection in Then

H = L2 (M )q

is a

M with q p. pxp p(x)p is a unitary from L2 (pM p) to pL2 (M )p which inter2 twines the left an right actions of pM p. Hence pM p on pL (M )q is unitarily 2 equivalent to pM p on L (pM p)q . So by (iv), dimpM p (pH) = trpM p (q ) = trM (p)1 trM (q ) = trM (p)1 dimM H. 2 2 Now if H is arbitrary, it is of the form e(L (M ) (N)) for e (M 1) / But e is the orthogonal sum of projections all equivalent to ones as in (iv) with q p. H = e(L2 (M ) 2 (N)) so M = e(JM J B ( 2 (N))e and p denes the isometry in the denition of dimM (pH). But p is a projection less than e in a II factor so by uniqueness of the trace, dimM (pH) = tr(M 1) (p) = tr(M 1) (p)/tr(M 1) (e) dimM (H) = trM (p) dimM (H).
(vi) We may suppose 61

L2 (M ), dimM (H) dimM (H) = 1 so by (v) and 2 (vi) the result is true for M -modules of the form L (M )p. Also if one forms k K = i=1 H then dimM 1 (K) = k dim H and dim(M 1) K = k 1 dimM by 2 2 k (v). But any H can be obtained from L (M ) as i=1 L (M )p for suitable k and p.
(vii) Observe that, on

Example 10.2.2

the unitary

If

0 < are icc groups, vN (0 ) acts on

( )
2

unitary between dimvN (0 ) ( 2 ())

(0 ) and = [ : 0 ].

of the right regular representation gives 2 (0 1 ). Hence by the coset decomposition,

(). And if a vN (0 )-linear

This inspires the following denition: If

[M : N ]

of

in

is the real number

N M are II1 dimN L2 (M ).


implies

factors, the index

Exercise 10.2.3 Example 10.2.4

Show that

[M : N ] = 1

N = M.

(Due to Atiyah and Schmidt.)

Discrete series representations of locally compact groups. Reduction by a nite projection in the commutant of a type II1 factor occurs in the representation theory of locally compact groups. If a discrete series representation is restricted to an icc lattice it generates a type II1 factor . The coupling constant is given by the ratio of the formal dimension and the covolume of the lattice. We illustrate in the case of

P SL(2, R)

which is the group of transfor-

mations of the upper half plane dened by invertible real is a fundamental domain illustrated below: DO FIGURE

H = {z C : Im(z ) > 0}, z a b c d

az + b cz + d

2 2 matrices D

It is well known that there

for the action of the subgroup

= P SL(2, Z)

P SL(2, Z) cover H and are disjoint apart from boundaries which are of Lebesgue measure 0. Thus if is an 2 invariant measure equivalent to Lebesgue measure, L (H, d) gives a unitary representation of which is unitarily equivalent to the left regular repre2 2 sentation tensored with the identity on L (D, d), making L (H, d) into a vN ()-module whose vN () dimension is innite. dxdy The measure is -invariant but we want to vary this procedure y2 dxdy slightly. For each n N consider . This measure is not invariant but y 2 n
The set and all its translates under 62

we can make the action of

P SL(2, R) unitary on L2 (H, f (z ) =

dxdy ) by the formula y 2n

a b c d

1 az + b ) f( n (cz + d) cz + d

(with perhaps an inverse matrix...exercise as usual). This changes noth-

P SL(2, Z) so we obtain (unitarily dxdy 2 equivalent) vN ()-modules Hn = L (H, ) for each n. y 2 n The commutant of vN () on Hn is a II factor. But as is well known, 2 holomorphic functions form a closed subspace of L functions which is manifestly invariant under P SL2 (R). The ensuing unitary representation is known to be irreducible and in the discrete series of P SL2 (R). It can be shown to be a nite projection in . Thus we have a concrete example of a vN ()-module with nite vN ()-dimension or coupling constant. In general, if G is a locally compact group with Haar measure dg , the
ing as far as how the representation looks to discrete series representations are precisely those irreducible unitary representations that are direct summands of the left regular representation on L2 (G, dg ). So if is a discrete subgroup with a fundamental domain D so that

is covered by the

(D)

which are disjoint up to measure zero sets,

we may apply the same analysis as above to obtain a

vN ()

module. The

obvious question is to calculate its coupling constant. This turns out to be quite simple because of a key property of discrete series representations. series representation

H is a Hilbert space aording a discrete G, then the functions g g , , the so-called coecients of are in L (G, dg ). We may then imitate the usual procedure 2 for nite or compact groups embedding H in L (G, dg ). And the usual Schur orthogonality of the coecients of a representation yields a number d such
See [ref robert] for the proof that if

of 2

that

d
G
If

g , , g dg = ,

, .

G is compact and Haar measure is normalized so that G has measure 1, d is the dimension of the vector space H. In general d depends on the choice of Haar measure but obviously the product of d with the covolume dg D 2 does not. The coecients give an explicit embedding of H in L (G, dg ) and
a straightforward calculation of the trace of the projection onto the image of

in

vN ()

yields immediately the formula

dimvN () (H) = d covolume().


The detailed calculation from this point of view can be found in [1] pp. 142148.

63

Example 10.2.5
bedding). a subspace of

Consider the subfactor

Each matrix unit eij gives an L2 (N Mn (C)) which is an orthogonal direct sum of these 2 equivalent subspaces. Hence [N Mn (C) : N ] = n .

N N Mn (C) (diagonal emN -linear isometry of L2 (N ) onto

Proposition 10.2.6
(a) (b)

If

is a type II1 factor on

then

M M

has a separating vector if has a cyclic vector if

dimM (H) 1.

dimM (H) 1. H
to

Proof. Both assertions follow immediately by comparing direct sum of copies of it.

L2 (M )p

or a

to control arbitrary vectors in on of

In fact both conditions in the last proposition are i. For that one needs L2 (M ). In fact the original denition of the Let

coupling constant by Murray and von Neumann was as follows.

be a type II1 factor whose commutant is a type II1 factor .

Choose

any nonzero vector

and let

and

and

respectively.

Then

q be p M

projections onto the closures and

q M

and using the

normalised traces the coupling constant was dened as the ratio the hard part being to show that this ratio is independent of constant with our

trM (q ) , trM (p)

Assuming

this last statement it is trivial to identify the Murray-von Neumann coupling

dimM (H) but at this .

stage we have nothing to oer in the

way of a simplied proof of why this number does not depend on

Example 10.2.7

(due to M. Rieel) If

(X, )

is a measure space and

is

a countable group acting by measure preserving transformations on (X, ) 2 so that acts by unitaries u on L (X, ) in the obvious way. We say that

D X is a fundamental domain for if X = (D) and (D (D )) = 0 for all , = id. (One may clearly suppose the (D) are disjoint by removing a set of measure zero.) In this situation the abelian von Neumann algebra L (X ) of -invariant L functions may be identied with the space L (D ). Now suppose and are two groups acting on X as above with fundamental domains D and E respectively. We may consider the von Neumann 2 algebra M, on L (X, ) dened as {{u : } L (X ) } .
a measurable subset

64

Chapter 11 The Crossed Product construction.


Perhaps the most useful way of producing von Neumann algebras from others is the crossed product. We begin by dening a very general notion about which there is not a lot to say, but then examine it carefully in special cases.

11.1 Group actions.


Let

be a von Neumann algebra and

a group. An action of

is a homomorphism

g g

from

to the automorphism group

G on M AutM of

(where automorphisms may be assumed ultraweakly continuous if necG essary). The algebra of xed points for the action is denoted M and is a von Neumann algebra . A special case of some importance is when the is a unitary group representation g ug with ug M ug = M g G. In that case setting g (x) = ug xug denes an action of G on M (and M ). We say that

is implemented by the unitary representation ug . If the ug are actually in M , we say that the action is inner as an inner automorphism of M is by denition one of the form Ad u(x) = uxu for u a unitary in M . An
the action automorphism is called outer if it is not inner. Actions are not always implementable though the notion depends on the Hilbert space on which

acts.

Exercise 11.1.1

T is a bijection of X 1 which preserves the measure class of (i.e. (A) = 0 (T (A)) = 0.) show how T denes an automorphism T of L (X, ). Show further that 2 this automorphism is implemented by a unitary u on L (X, ).
If is a measure space and
A bijection subset

(X, )

as above is called ergodic if

AX

implies either

(A) = 0

or

T (A) = A (X \ A) = 0.

for a measurable

65

Proposition 11.1.2
points for

With notation as above

is ergodic i the only xed

are constant functions.

f L and T (f ) = f . After throwing away a null set we may assume that f (x) = f (T (x)) for all x X . Then for every > 0, by the denition of the essential supremum, ({x : ||f || |f (x)| < } = 0. But this set is invariant under T so it is equal to X up to a set of measure 0. Letting tend to 0 we see that ({x : |f (x)| = ||f ||}) = 0. So we may assume f (x) = eig(x) for some measurable g taking values in [0, 2 ). Repeating the argument for g gives f constant almost everywhere. () If A is a measurable invariant set then its characteristic function is xed by in L i A is invariant.
Proof. () Let

Exercise 11.1.3
implementable for

Let

x =

0 1 1 0

y =

be the Pauli spin matrices. action of the group

Show that

0 i i 0 Ad ux , Ad uy

and

z =

1 0 0 1

and Ad uz dene an

Z/2Z Z/2Z M2 (C) on C2 .

on the two by two matrices which is not

Exercise 11.1.4

Show that any group action is implementable for a type

II1 factor in standard form and more generally any automorphism group pre-

serving a faithful normal state is implementable in the GNS representation.

Exercise 11.1.5 Exercise 11.1.6


If

Show that every automorphism of Show that the automorphism of

B (H)

is inner. coming from the

vN (F2 )

group automorphism which exchanges the 2 generators is outer.

is a topological group there are many possible notions of continuity.

The most useful is that of pointwise *-strong convergence, i.e. we assume that the map

g (g )(x)

is *-strong continuous for any

x M.

Typically many

other notions of continuity will be equivalent to that and even a measurability assumption can be enough to ensure this continuity. We will always assume pointwise *-strong continuity when referring to an action of a topological group.

Exercise 11.1.7

Is the action by translation of

R on L (R) pointwise norm

continuous? pointwise strongly continuous? pointwise *-strong continuous?


Actions of a given group on von Neumann algebras are easy to construct but actions of a group on a given von Neumann algebra may be hard to come by.

66

Denition 11.1.8 Exercise 11.1.9


sulting action of

An action of

on

is said to be ergodic if

M G = Cid.

Show that if G acts preserving on (X, ) then the re on L (X, ) is ergodic i the only measurable subsets satisfy either

A X which satisfy (g (A)A) = 0 g G (X \ A) = 0. (Here AB means A \ B B \ A.)

(A) = 0

or

The following question is an intriguing open problem: Does

SU (3)

have any ergodic action on a type II1 factor ?

It is shown in [] that

SU (2)

has no such action and it is shown in [] that

if a compact group acts ergodically on a von Neumann algebra then that von Neumann algebra has a faithful normal trace.

11.2 The crossed product


If

is an action of the locally compact group G with Haar measure dg on the von Neumann algebra M on the Hilbert space H. Form the Hilbert space K = L2 (G, H) = L2 (X ) H and let G act on K by ug = g 1, being the left regular representation. Further, let M act on K by ( xf )(g ) = g1 (f (g ))
.

Exercise 11.2.1
of

Show that

onto a von Neumann

xx is an ultraweakly continuous *-isomorphism subalgebra of B (K). ug x u g = g (x).

Exercise 11.2.2

Show that

Note that this gives another way of making a group action implementable, at least when it is locally compact.

Denition 11.2.3
M G is the {x : x M }.

If

M , H, G

and

are as above, the crossed product

von Neumann algebra on

generated by

{ug : g G} M
with

and

. Note M that nite linear combinations g xg ug form a dense *-subalgebra of M G. Moreover the ug are linearly independent over M in the sense that g x g ug = 0 xg = 0 for each g in the sum. This dense subalgebra could be called the
>From now on we will drop the

and identify

algebraic crossed product.


67

G is compact or abelian, but we shall be mostly interested in the case where G is discrete as then we may replay the matrix element game that we played for vN () to gain control
There is a well-developed theory of of weak limits of elements in the algebraic crossed product. (In fact of course

G when

vN () is the special case of the crossed product when M = C and the action
is trivial.) Indeed we see

G is discrete, any element of M G g xg so that the sum g xg ug stands for a certain 2 matrix of operators on K = H (G). Moreover any matrix of this form which denes a bounded operator on K is in M G. This is because the sum converges pointwise at least on the dense set of functions of nite support from G to H. In the case where the
immediately as in 3.1.10 that if denes a function crossed product is a II1 factor we know that the commutant consists of right multiplication by elements of right multiplication by case later on. Moreover the formulae

G so a wealy dense subalgera of

preserves this dense subspace of vectors and on that subspace

(M G) g xg ug and

ug

and

xM

commute. We will return to the general

(
and

xg ug ) = y g ug ) =
g

g (xg1 )ug {
h

xg ug )(

xh h (yh1 g )}ug

are justied by matrix multiplication. We shall now provide some sucient conditions for always assuming

G to be a factor

is discrete.

Denition 11.2.4
G
for which

An action

of

on

is called outer if the only

in

g M

is inner is the identity. If

Proposition 11.2.5
on the factor
Proof. If that next

is a discrete group and


G is a factor.

is an outer action of

then

xg ug Z (M ) then equating coecients in the expression x commutes with M gives us yxg = xg g (y ) y M ,g G. By the lemma this implies xg = 0 for any g = 1. Thus x M and since M is x=
Let

a factor we are done.

Lemma 11.2.6
x = 0,
Then with

Aut M

for a factor

M.

Suppose there is an

x M,

yx = x(y )
is inner.
68

y M.

were unitary this would be obvious. So take the adjoint of the relation to obtain x y = (y )x y M . Thus yxx = x(y )x = xx y and xx Z (M ). Similarly x x Z (M ). But xx and x x always have the same spectrum so since M is a factor both xx and x x are equal to the same positive number are done. These two results prompt the following denition.

Proof. If

Dividing by

converts

into a unitary and we

Denition 11.2.7
called free if An action

An automorphism

of a von Neumann algebra

is

yx = x(y )
is called free if

y M x = 0.
is free for every

g = id.

The argument of proposition 11.2.5 shows in fact that if on a von Neumann algebra

then

is a free action Z (M G) M , in fact that M G

M. is a free ergodic M G is a factor.


If action of on a von Neumann al-

Theorem 11.2.8
gebra

M,

then

Proof. This follows immediately from the preceding remark. To understand the meaning of freeness for automorphisms of the form

T we need to make a hypothesis on (X, ) as otherwise one could envisage a T which is non-trivial on X but for which T is the identity. So we will suppose from now on that (X, ) is countably separated. This means there is a sequence Bn of measurable sets with (Bn ) > 0 for which, if x = y , there is an n with x Bn but y / Bn . Obviously Rn is countably separated.

Exercise 11.2.9

Show that

T = id

means that

Tx = x

almost everywhere.

Hint-look at the proof of the next result.

Proposition 11.2.10
({x : T (x) = x}) = 0.

If

is a transformation of

(X, )

then

is free i

Proof. ()If A is any measurable set on which for all f L .

T = id

then

A f = T (f )A

() First throw away any xed points of T . Then suppose f1 T (f2 ) = f2 f1 f2 L . Let A be the support of f1 . Then since T has no xed 1 points, A = n (A Bn \ T (Bn )). If f1 were non-zero in L , we could thus choose an n for which (A Bn \ 1 T (Bn )) > 0. Set f2 = Bn . Then for any x A Bn \ T 1 (Bn ) we have 69

f1 (x)f2 (x) = 0 but f1 (x)f2 (T x) = f1 (x)Bn (T x) = 0 since x / T 1 (Bn ). Thus f1 T (f2 ) = f2 f1 in L . So the measure of A must be zero. We conclude that if is a countable group acting freely and ergodically on a measure space (X, ), preserving the class of , then the crossed product L (X, ) is a factor. Note that if is abelian, ergodic implies free.

Exercise 11.2.11

Show that freeness of the action actually proves that

L (X, )

is maximal abelian in the crossed product.


The crossed product

when

is abelian and

is discrete is called

the group measure space construction. Here are several examples.

Example 11.2.12 X = Z, = Z
measure. The action is free and ergodic and

acting by translation,

counting

L (X, ) = B ( 2 (Z)).

Example 11.2.13
sion.

The irrational rotation algebra-von Neumann algebra verwhere

(X, ) = (T1 , d), = Z generated by the transformation T e z and /2 is irrational.


i

T (z ) =

Exercise 11.2.14 Example 11.2.15


Then

Use Fourier series to show that this


Let

is ergodic.

be a nite abelian group and

the countable group of sequences

(hn )

nN H be with hn eventually the identity. Put

X = G = G is a

nN H (the set of all sequences) with the product topology. compact group so has a Haar measure . acts on X by left

translation. The action is clearly free and ergodic as we shall now argue. There is a particularly von Neumann algebraic way to view this example without even constructing the space (X, ) ! be the group algebra of the dual group Let A = L (H ) = CH usual trace. As in section 7.2, form the algebraic tensor with product trace

, with its H product alg,nN A

tr.

Then perform the GNS construction with respect to

tr to obtain an abelian von Neumann algebra. It may be identied with L (G, ) so the Hilbert space H of the GNS construction is L2 (X, ). But it is clear that an orthornormal basis of H is given by nite sequences (n ) of which dene elements 1 2 1 1 1 in alg,nN A. elements of H The point is that these basis vectors are eigenvectors for the action of on L2 (X, ): (hn )(1 2 1 ) = (
n
70

n (hn )) 1 2 1 .

Ergodicity follows easily since the only basis element which is xed by all the

(hn )

is the one with all

equal to

1.

Exercise 11.2.16

H = Z/2Z in this example then the subalgebra of the crossed product generated by alg,nN A and is the algebraic innite tensor product of M2 (C).
Show that if
Both of the last two examples are special cases of a more general one:

is a compact group with its Haar measure and

is a countable dense

subgroup acting (freely) by left translation. The Peter Weyl theorem shows that this action is ergodic.

Example 11.2.17
If

Bernoulli Shift.

is any innite group and A = CZ/2Z we may form the tensor product indexed by of a copy of A for each . The von Neumann algebra thus obtained is once again the L space of the innite product measure space, this time with the set indexing the product being . As in the previous 2 example we can obtain a basis of L indexed by funcions from to the set {0, 1} which are almost always 0. These functions are the same as nite subsets of and the action of on the Hilbert space is by permuting the
basis in the obvious way. Ergodicity follows from the fact that the orbit of any non-empty subset is innite. One could also chose another trace than the usual one and modify the orthonormal basis of specied. We give a few more examples of free ergodic actions without supplying proofs of ergodicity.

A accordingly.

The measures are the obvious ones unless

Example 11.2.18 SL(2, Z) acts on T2 = R2 /Z2 via the linear action on R2 . Example 11.2.19 P SL(2, Z) acts on R {} by linear fractional transformations.

Example 11.2.20 SL(2, Z) acts on R2 by linear transformations. Example 11.2.21 Q acts on R by translation.
There are two fairly easy ways to see that this action is ergodic. The rst is to reduce it to a dense subgroup of a compact group by observing that an L function on R which is invariant under translation by Z denes an L function on the quotient

T.Then

use Fourier series.

The second way is a direct attack which should generalise to show that bullshit 71

translation by any countable dense subgroup of a locally compact group is ergodic. If f L (R) is invariant under Q, set things up so that there are sets and

A B

and

both of nonzero measure, so that

g (A) g (B ) = .

Cover

with intervals of the same width with rational endpoints.

Some of

these must intersect

and

translates of each other so

B in non-nul sets. But all these intervals are all g cannot be invariant up to sets of measure zero. Q Q
acts on

Example 11.2.22 Example 11.2.23


normalised trace on

The ax+b" group

R
but using a

Same as example 11.2.13 with

H = Z/2Z CH

CH

which is dierent from the usual one. Such a trace which we could call The product measure is not absolutely continous

is specied by its values on the minimal projections of

and

1p

for

0 < p < 1.

with respect to Haar measure, and it is not preserved by group translation so this example is perhaps most easily approached by the von Neumann algebra construction where one can implement the action of nN Z/2Z by 2 unitaries. These unitaries come from ones on L (H ) which exchange two points of unequal weight so they must be correctly scaled.

Exercise 11.2.24
Type I : Type Type

Work out the details of example 11.2.23

In the examples we see four dierent kinds of free ergodic actions:

acts transitively.11.2.12

Type II1 :

preserves a nite measure. 11.2.13,11.2.15,11.2.17,11.2.18 II : preserves an innite measure.11.2.20,11.2.21 III : preserves no measure equivalent to .11.2.19,11.2.22,11.2.23

11.3 The type of the crossed product.


We adopt the notations and conventions of the previous section. The map

Em : M

M which assigns aid to the element is destined to play a big role in the theory. It is called the conditional expectation onto M and
obviously satises the following contitions:

2 EM = EM . (ii) EM (x) = EM (x ), EM (1) = 1, EM (x x) = 0if f x = 0. (iii) EM (x x) EM (x )EM (x), ||E (x)|| ||x||. (iv) EM (axb) = aEM (x)b for a, b M . (v) EM is ultraweakly continuous.
(i) So

EM

is a projection of norm one in the Banach space sense.

The

condition (iv) says that

EM

is an

M M -bimodule
72

map.

acts non-transitively,freely and ergodically, preserving the nite measure then L (X, ) is a II1 factor. If preserves the innite -nite measure then L (X, ) is a II factor unless acts transitively in which case L (X, ) is type I.
If
Proof. (i) It is clearer to prove a more general statement (in the case where preserves

Theorem 11.3.1

and

(X ) = 1).

So suppose

preserves the faithful positive

ultraweakly continuous trace action is free and ergodic.

tr

on the von Neumann algebra

and that its

Then we claim

M = A

is a type II1 factor

(or a nite dimensional factor). By previous results we need only show that it has an ultraweakly continous positive trace. So dene

T r = tr EA

on

M.

Ultraweak continuity and positivity are obvious so by continuity and

T r(au bu ) = T r(bu au ). For either side of 1 the equation to be non-zero means = amd then the left hand side is 1 1 tr(a (b)) = tr( (a (b))) = tr(b (a) ) which is equal to T r(bu au ). (ii) If is innite and does not act transitively then there are no atoms hence there are subsets Y of X of arbitrary positive measure. Let Y have nite non-zero measure and let be the function ( ) = ,id Y . Then au , = (au ) = id, Y a(x)d(x). One easily checks that ((pau p)(pbu p)) = ((pbu p)(pau p)) so by 4.0.25 denes a positive ultraweakly continuous trace on p(A )p which is a type II1 factor . But A is not itself a type II1 factor since A contains
linearity it suces to prove an innite family of equivalent mutually orthogonal projections. By 9.1.8 we are done. (iii) If

acts transitively then

(X, ) = (, counting

measure) and the

characteristic function of a set with one element is a minimal projection in L (X, ) .

Exercise 11.3.2
sure

If

acts ergodically on

(X, )

preserving the

then any other invariant equivalent measure is proportional to

-nite mea.

We now want to show that there are factors that are neither of type I nor type II . Suppose that M = L (X, ) is a type I or II factor. Then it has a trace absolutely continous with respect to

X, ,by reversing the procedure of theorem 11.3.1 and dening the measure (A) to be tr (A ) (A L (X, ) M ). Invariance of the measure is no problem. The snag is that tr (A ) could be innite for every non-nul set A. We will show that this is not the case. To
We would like to dene an invariant measure on this end the concept of lower semicontinuity will be useful.

tr : M+ [0, ].

73

Denition 11.3.3
such that

If

is a topological space we say that for all

lower semicontinous if for every

f (u) > f (x) ((, K ]))


If

x X and > 0 u U. K R.

there is an

f : X R is open set U X

Exercise 11.3.4
(a)f
1

Prove that if

is lower semicontinous then

is closed for every

(b)f attains its minimum on any compact subset of

X. a ||a ||

Exercise 11.3.5
from

H f

is a Hilbert space and

H,

the function

B (H)

to

is weakly lower semicontinuous. If are lower semicontinous then

Exercise 11.3.6
tinous if it exists.

is lower semicon-

Lemma 11.3.7

Let

be a type I or II factor and

tr : M + [0, ]

be

T race in type I,as in 9.1.9 in type II each K 0, M1,K = {x : tr (x x) K }


with that

and the trace in type II1 . Then for is weakly compact.

Proof. Clear in the II1 case. In a decomposition

M = N B( 2 (N))

on

N a type II1 factor or C we may assume by 10.2.6 that there is a vector e11 H with a trace on e11 M e11 . So if i = ei1 we have, up to a scalar,

tr(x) =
i=1

xi , i .

By the previous exercises and weak compactness of the unit ball, we are done.

x M1,K let W (x) be the weak closure of the convex set of nite sums { i i = 1, i > i i ui xui : 0, ui unitary in L (X, )}. Then W (x) M1,K and if (y ) = tr(y y ) for y W (x) then attains its minimum at a unique point E (x) of W (x).
With notation as above, for
Proof. Note rst that {z M : tr (z z ) < } is a vector space on which ||z || = tr(z z ) denes a prehilbert space structure. (Since (a + b) (a + b) 2(a a + b b) as operators, and the parallelogram identity passes to the potentially innite sum dening of

Proposition 11.3.8

tr.)

Moreover

W (x)

is a weakly compact subset

so by lower semicontinuity

attains its minimum at a point which is

unique by two dimensional Euclidean geometry as in 2.1.2.

Proposition 11.3.9
projection in

Suppose that

for a free ergodic action of

with

on tr(p) < . Then

M = L (X, ) is a type I or II factor L (X, ). Let tr be as above and p be a

E (p) = EL (X,) (p)


and

0 < tr(E (p)2 ) tr(p).


74

Proof. Let

E = EL (X,) . By the uniqueness of E (p) it commutes with every unitary in L so it is in L by 11.2.11. On the other hand E (y ) = E (p) for all y W (p) by the bimodule linearity of the conditional expectation and its ultraweak continuity. So E (E (p)) = E (p) = E (p). But (E (p) (p) = tr(p). Finally E (p) = E (p2 ) which is a positive non-zero selfadjoint operator and hence has non-zero trace.

Theorem 11.3.10
rated measure on

Let

act freely and ergodically on the countably sepa-

-nite measure space (X, ) so that there is no -nite -invariant X absolutely continuous with respect to . Then L (X, ) is

a factor not of type I or II.


Proof. If the crossed product were of type I or II, dene the measure

on

By the previous result (A) would have to be nite 2 and non-zero for some A since the L functionf = E (p) must dominate a by multiple of to

(A) = tr(A ).

A for some A (e.g. let A be those x with f (x) suciently close ||f ||). But then by ergodicity X = (A) (up to null sets) so that is -nite. It is automatically absolutely continuous wrt . Invariance of 1 under follows from tr (u xu ) = tr (x) for x 0.
A factor not of type I or II is called a type III factor.

Denition 11.3.11

Example 11.2.22 provides a type III factor since the subgroup ergodically so the only possible invariant measure is a multiple exercise 11.3.2 and this is not invariant under multiplication!

Q acts of dx by

Note that the above technique works in somewhat greater generality than actions of groups on measure spaces.

Exercise 11.3.12
M

Adapt the proofs of the results just obtained to show that

Z is a type III factor if the action

is generated by a single automor-

phism of the II factor scaling the trace by a factor

= 1.

75

76

Chapter 12 Unbounded Operators and Spectral Theory


There are many naturally arising examples of unbounded operators, some of the most fundamental being multiplication by

and dierentiation, the

position and momentum operators of quantum mechanics. Our immediate motivation for studying unbounded operators here is to facilitate the study of arbitrary von Neumann algebras acting on GNS Hilbert spaces. Here we establish the necessary preliminaries on unbounded operators. The material closely follows Reed and Simon [2].

12.1 Unbounded Operators


Denition 12.1.1
subspace An operator

on a Hilbert space

consists of a linear

D(T ),

the domain of

T,

and a linear map from

D(T )

to

H.

Example 12.1.2
(i)

Mx ,

multiplication by

on

L2 (R). x2 |f (x)|2 d x < .


R

D(Mx ) =

f L2 (R) :

(ii)

T =

d on dx

L2 ([0, 1]). D(T ) = C 1 [0, 1].

In order to do some analysis we want to restrict our attention a little so as not to consider completely arbitrary linear maps.

77

Denition 12.1.3
T (T )

Let

be an operator on

H.

The graph of

is

(T ) = {(, T ) : D(T )} H H.
is closed if is closed in

H H.
is closed and

Remark 12.1.4 Lemma 12.1.5


(0, )

Note that if

D(T ) = H

then

is bounded

by the Closed Graph Theorem. A linear subspace

HH

is the graph of an operator i

implies

= 0.

Proof. Straightforward. Many operators are not closed, but can be extended to a closed operator.

Denition 12.1.6
denoted

S T,

if

Let S, T be (S ) (T ). An operator

operators on Equivalently

H. T is an D(S ) D(T )

extension of

and

S, T |D(S ) = S .

Denition 12.1.7
extension.

is preclosed (or closable) if it has a closed

Lemma 12.1.8
tension

Suppose

is preclosed. Then

has a smallest closed ex-

T . (T ) = (T ).
is called the closure of

Denition 12.1.9 T
Proof of Lemma. contains because:

T. A
of

Take a closed extension

T . (A)

is closed and

(T )

so

(T ) (A). (T )

is the graph of an operator (call it

T)

(0, ) (T ) (A) = A(0) = 0. T is the smallest closed extension because for all closed extensions A, (T ) = (T ) (A).

Remark 12.1.10
operator: (i)

We thus obtain two equivalent denitions of a preclosed

(0, ) (T ) = 0. D(T ), n 0
and

(ii) (n

T n

converges)

T n 0.

Exercise 12.1.11
(i) Dene

on

L2 (R)

with compact

D(S ) = C0 (R) (innitely dierentiable support), Sf = f . Show that S is preclosed.

by

functions

78

(ii) Dene

from

L2 (R)

to

by

D(T ) = L1 (R) L2 (R), T (f ) =

Show that

is not preclosed. Suppose

f.

Denition 12.1.12
subspace

D0 D(T )

such that

T is a closed T |D0 = T .

operator. A core for

is a linear

We can perform basic arithmetic operations with (unbounded) operators as follows: S + T is the operator with domain D (S + T ) = D (S ) D (T ) and (S + T ) = S + T . ST is the operator with domain D(ST ) = { D(T ) : T D(S )} and (ST ) = S (T ). Of particular importance is the adjoint.

Denition 12.1.13

Let

be a densely dened operator on

H.

Let

D(T ) = { H : H such that < T , >=< , > D(T )} = { H : C > 0 such that | < T , > | C || || D(T )}.
For and

D(T ) note dene T = .

that the element

is unique (by the density of

D(T ))

Remark 12.1.14 Exercise 12.1.15

Note that if

ST

then

T S .

Give an example to show that the domain of the adjoint

need not be dense. [In fact it can be

{0}].

Proposition 12.1.16
1. 2.

Let

be a densely dened operator. Then

is closed. is dense i

D(T ) T

is preclosed. In that case

T = T .

3. If

is preclosed then

(T ) = T .
i

Proof. Note that

(, ) (T ) < (T , ), (, ) >= 0. Hence

< T , >=< , >

for all

D(T )

i

(T ) = {(T , ) : D(T )} = (u(T )) = u(T ) ,


where

u:HHHH

is the unitary operator

u(, ) = (, ).
is closed.

Now:

1. Orthogonal complements are closed, hence

(T )

79

2.

(T ) = ((T ) ) = u (T ) ,

so

(0, ) (T ) (, 0) (T ) 0 =< (, 0), (, T ) >= < , > for D(T ) .


Hence

all

D(T )

is preclosed i

D(T ) = {0}
2

i

D(T )

is dense. so

In that case

(T ) = u(T )

= u (T )

= (T ) = (T ),

T = T .
3.

T = T = T = (T ) .
An operator

Denition 12.1.17

T is symmetric if T T . Equivalently < T , >=< , T > for all , D(T ). T is self-adjoint if T = T . A self-adjoint operator T is positive if < T , > 0 for all D (T ).

12.2 Spectral Theory for Unbounded Operators


Denition 12.2.1
Let

be a closed operator on

H.

The resolvent of

is

(T ) = {|1 T : D(T ) H
The spectrum of

is a bijection}.

is

(T ) = C\(T ). 1 T : D(T ) H
is a bijection then

T )1

Remark 12.2.2 Exercise 12.2.3

Note that if

(1

is bounded by the Closed Graph Theorem. The spectrum is highly dependent on the domain. set of absolutely continuous functions on Let

AC [0, 1] denote the d d , T2 = dx , with dx

[0.1].

Let

T1 =

D(T1 ) = {f AC [0, 1] : f L2 ([0, 1])} D(T2 ) = {f AC [0, 1] : f L2 ([0, 1]), f (0) = 0}.
Show that

T1

and

T2

are closed. Show that Let

(T1 ) = C

while

(T2 ) = .

F a measure2 able, real-valued, a.e. nite function on X . Let D (Mf ) = {g L (X, ) : 2 f g L (X, )} and let Mf g = f g . Then Mf is self-adjoint and (Mf ) = ess.range(f ) = { C : ({x : | f (x)| < }) > 0 > 0}.
be a nite measure space and
80

Proposition 12.2.4

(X, )

Exercise 12.2.5

Prove Prop 12.2.4. Let

Theorem 12.2.6 (Spectral Theorem - Multiplier Form)


adfoint operator on space

A be a self-

H with dense domain. Then there exists a nite measure (X, ), a real-valued a.e. nite function f on X and a unitary operator u : H L2 (X, ) such that uAu = Mf
Proof. See [2]

Remark 12.2.7 (Borel Functional Calculus)

Note that the Spectral The-

orem allows us to dene a Borel functional calculus for self adjoint operators. Given a Borel function h on the spectrum of A, dene h(A) = u Mhf u.

12.3 Polar Decomposition


Theorem 12.3.1
Then: (i) Let

A : H K

be a closed, densely dened operator.

A A and AA are (AA )1/2 exist).


nal space

positive self-adjoint operators (hence

(A A)1/2

and

(ii) There exists a partial isometry with initial space

Range(A A)1/2

and

Range(A)

and

A = v (A A)1/2 .
(iii) If

A = uB for some positive B and partial Range(B ) then u = v and B = (A A)1/2 . A = (AA )1/2 v .

isometry

with initial space

(iv) In addition
Proof. (i) Since

P : (A) H be projection onto the rst component. Since A is an operator Ker(P ) = {0} and hence Range(P ) is dense in (A) (so P P H is a core (A)
is closed, it is a Hilbert space. Let for

A).

Let

H, P = (, A ).

Then, for all

D(A),

< , >=< P , (, A ) >=< , > + < A, A > < , >=< A, A > A D(A ) and A A = .
81

D(A A) P P H which is a core for A. 1) = H.


Thus

In addition

Range(A A +

It is easy to see that A A is symmetric, so A A + 1 (A A + 1) . Let D ((A A + 1) ). Since Range(A A + 1) = H there exists ). D(A A + 1) with (A A + 1) = (A A + 1) (= (A A + 1) Ker((A A + 1) ) = {0} because Range(A A + 1) = H, and hence D(A A +1). Thus (A A +1) = A A +1 and so (A A) = A A. =
Finally, for positive,

D(A A), < A A, >=< A, A > 0 so A A i.e. (A A) [0, ) (just use the Spectral Theorem).

is

(ii) As we noted above, D (A A) is a core for A. D (A A) is also a core for 1/2 |A| = (A A) (use spectral theory). Thus AD(A A) = RangeA and |A|D(A A) = Range|A|. Note that for D(A A),

|||A| ||2 =< A A, >=< A, A >= ||A ||2 , v : |A| A , D(A A), extends to a partial isometry with initial space |A|D (A A) = Range|A| and nal space AD(A A) = RangeA.
so that the map For

D(|A|) take n D(A A) An = v |A|n v |A| and, as A is D(A) |A|


take

(n , |A|n ) (, |A| ). Then closed, D (A) and A = v |A| .


with

For

n D(A A)

with

(n , An ) (, A ).

Then

|A|n = v v |A|n = v An v A.
Since is closed,

D(|A|).
and

Hence (iii) If

D(A) = D(|A|)
then

A = v |A|.

A = B u = Bu . A A = Bu uB = B 2 since u u is projection onto Range(B ). By uniqueness of the positive square root 1/2 of a positive operator (Exercise 12.3.3), (A A) = B . Thus the initial space of u is Range(|A|) and u|A| = A = v |A| so u = v . A = uB

(iv)

A = v (A A)1/2 so A = (A A)1/2 v and hence AA = v (A A)1/2 (A A)1/2 v = v (A A)v (Exercise 12.3.3). Thus v implements the unitary equivalence 1/2 = v (A A)1/2 v and of AA |Range(A) and A A|Range(A ) . Hence (AA ) 1/2 then A = v (A A) = (AA )1/2 v .
Note that it was very important in (i) to establish that contained a core for A and hence was dense. It was not clear a D(A A) contained any elements other than 0. (i) Let

Remark 12.3.2
D(A A)

priori that

Exercise 12.3.3

be a positive operator. Show that

T 1/2 T 1/ 2 = T .

(ii) Show that a positive operator has a unique positive square-root.

82

12.4 Unbounded operators aliated with a von Neumann algebra .


If

M is a von Neumann algebra on H, an element a B (H) au = ua for every unitary in M . This inspires the following.
If

is in

i

T : D(T ) H is a linear operator on the Hilbert space H and M is a von Neumann algebra on H we say that T is aliated with M , written T M if, for any unitary u M , uD(T ) = D(T )
and

Denition 12.4.1

uT = T u D(T ).

Lemma 12.4.2

If

is preclosed with closure

then

T M

if

T M . M
. But

Proof. It is clear that

T M

i

u(T ) = (T )

for all unitaries in

this property passes to the closure of the graph.

Lemma 12.4.3
2. If

If

is a closed operator aliated with

then

1. The projection onto

(T )

is a

22

matrix of operators in

M. f (|T |) M

T = u|T |

is the polar decomposition of

then

uM

and

for any bounded Borel function of


Proof.

|T |.

1. is obvious from the characterisation of aliation given in the proof of the previous lemma. 2. follows from uniqueness of the polar decomposition and the bicommutant theorem.

83

84

Chapter 13 Tomita-Takesaki theory.


In chapter 9 we showed that the GNS construction on to

using a faithful

Htr with respect M and its commutant. This is because the map J , which is the extension to Htr of the * operation on M , is an isometry. So x JxJ is the extension to Htr of right multiplication by x . Unfortunately if we use a (normal) non-tracial state the * operation is no longer an isometry and there is no reason to expect either it or right multiplication by elements of M to have bounded extensions to H . But as we shall see, the * operation is actually 1/2 preclosed in the sense of unbounded operators and if S = J is the polar
normal trace produces a perfectly symmetric Hilbert space decomposition of its closure S , we will show that 1/2 it it will satisfy M = dynamics-a one parameter

JM J = M ! Quite remarkably, the operator M so that a state actually gives rise to a automorphism group of M (and M )!!

We will prove these results using a method of van Daele for which we have followed some notes of Haagerup ([],[]). But before getting started on this dicult theory it is essential to do some elementary calculations to see how it all works out in the

22

matrices.

Exercise 13.0.4
form i

Let M be M2 (C). Show that any state on M is of the (x) = T race(hx) for some positive h of trace 1. And that is faithful

is invertible. Thus with respect to the right basis,


1 1+

(x) = T race(x
for some

0
1+

, 0 1.
With notation as in the previous exercise, suppose be the * operation on the GNS Hilbert space
85

Exercise 13.0.5
faithful and let

is

H .

Calculate

1/2 the polar decomposition S = J and show that z z Show that M = M for z C so that z (x) =

SM S = JM J = M . z xz = M denes a

representation of

as automorphisms of

which are

-automorphisms i

z iR .

Exercise 13.0.6

Generalize the above to the

nn

matrices and in fact any

nite dimensional von Neumann algebra .

13.1 S,F and their graphs.


Throughout this section

will be a von Neumann algebra on

and

a cyclic and separating vector for

M and hence M . (The same data as a faithful normal state.) Let S0 and F0 be the conjugate linear operators with domains M and M dened by S0 (x) = x and F0 (x) = x for x M and M respectively.
In the sense of unbounded operators
F0 S0

Lemma 13.1.1
so that
Proof.

and

S0 F0

S0

and

F0

have densely dened adjoints and hence are preclosed.


S0 (a )
is dened if

conjugate linear map on all of in

S0 (a), a extends to a bounded S0 (a), a = (a ) , a which is bounded as a function of a by Cauchy-Schwartz. Hence a is the domain of S0 and S0 (a ) = (a ) = F0 (a ). Interchanging S0 and
To show

H.

But

F0
Let

we get the other inclusion.

Denition 13.1.2
S = J
1/2
Observe that

Let

and

F S

be the closures of

S0

and

F0

respectively.

be the polar decomposition of


1 S0 = S 0
Thus so

S. S2 = 1 2 kernel J = 1
in the sense of 1 /2 and J J =

is injective and has zero

unbounded operators.

1/2

1/2 .

The same goes for see that F = S .

and its polar decomposition, but we shall now

Theorem 13.1.3
set of all

(Takesaki,[].)

S = F , F = S

and the graph of

is the

and

(c, c ) where c D(c) D(c ).

is a closed densely dened operator aliated with

Proof. Let

(, F ) be in the graph of F . By the denition of F we know that , (a ) = a , F . Now dene operators a and b with domain M by ax = x and bx = x F . Then a and b are closable for if x and y are in M we have a(x ), y = x , y = , (x ) y
86

= (y ) x , F = x , y F = x , b(y ) b a . Let c be the closure of a. Then c = a = and c = a b so c = F . Now by construction the domain of a is invariant under the unitary group of M and on it a commutes with the unitaries in M . This means that c is aliated with M . At this stage we have shown that the graph of F consists of all (c, c ) where c is a closed densely dened operator aliated with M and D (c) D (c ). We now want to show that the graph of F is contained in the graph of S . c c be its polar decomposition. This is not hard. Let c be as above and c = Then if fn (t) = t for 0 t n and fn (t) = 0 for t > n we have that fn ( c c) c c on any vector in the domain of c, and since c is aliated with M , fn ( c c) M so that ufn ( c c) is in the domain of S and tends to . Moreover fn ( c c)u tends to c = F so (, F ) is in the graph of S . Thus F S and we have already observed that S F . Hence S = F and S = F .
so that as before and

a b

Corollary 13.1.4

The polar decomposition of

is

J 1/2 . M
.

We now prove a crucial result connecting

and

Lemma 13.1.5
with

in the

R+ be given. Then for a M 1 domain of F and a = (S + F )a.


Let

there is an

aM

Proof. Assuming

(x) = x, a
with

||a || 1 we may apply theorem 8.3.1 to the and (x) = x, to obtain the existence of x, a = ax, + 1 xa, = x, a + 1 a, x .

dened by an

aM

Provided

a is in the domain of F

this equation reads

a = (S + 1 F )a.

On the other hand rearranging the equation gives

a, x = x, a a
so by Cauchy Schwartz

is in the domain of

F = S .

Corollary 13.1.6
an M
with

an

For each , 1/2 an

D(1/2 ) D(1/2 ) there is a sequence 1/2 and 1/2 an 1/2 .


87

Proof. Set

= (S + F )

and choose a sequence

the previous lemma there are an M with F = J (1/2 + 1/2 ) has bounded inverse (in the usual sense of unbounded 1 1 operators) so put n = (S + F ) (an ). So an = (S + F ) an . Moreover 1/2 an = 1/2 (1/2 + 1/2 )1 Jan

an M with an . By (S + F )an = an . But S +

1/2 and (1/2 1/2 1

+ 1/2 )1 is bounded by spectral theory. So 1/2 an (S + F ) (S + F ) = 1/2 . In the same way 1/2 an 1/2 . M
and

We put everything together with a lemma linking subspace to which many functions of

on a dense

can be applied.

Lemma 13.1.7
then

If

and

are in

D(S ) D(F ), a ,

and

as in 13.1.5,

SaS, + 1 F aF , = a , .
Proof. By moving one

and

to the other side of the inner products, we

are x and y respectively, both in D (F ), for x and y in M . But on M , SaS acts by right multiplication by a so SaS, = xa , y = Sa, x y . On the other hand, systematically using F = S we obtain F aF x, y = y x, a = Sx y , a = F a, x y . Combining these two we see
see by the previous lemma that we may assume and

SaS, + 1 F aF , = (Sa + 1 F a), x y .


But by 13.1.5 this is

a , x y = a ,

13.2 Proof of the main theorem.


We begin with an easy exercise in countour integration.

Exercise 13.2.1
f

Let

be the strip

{z C : 1/2

is continuous and bounded on

and analytic on the

(z ) 1/2}. Suppose interior of S . Then

f (0) =

f (1/2 + it) + f (1/2 + it) dt 2cosht S

Hint: Integrate

tending to the boundary of

f (z ) sin z

around rectangular contours in

S.
88

Proposition 13.2.2

With notation as in the previous section

a=

Proof. Since

2it

it Ja J it dt 2cosht J (D(S ) D(T )) = D(S ) D(T )


so

1/2

J =

1 /2

we have

after a little rearrangement the formula of 13.1.7 reads

Ja J, = a1/2 , 1/2 + 1 a1/2 , 1/2 .


Now let all

H0 be the dense subspace of all vectors in H which is the union of [a,b] ( for 0 < a < b < . Certainly H0 D(S ) D(F ), H0 is invariant z z under J and for z C, and moreover for H0 , z is an entire function of z .
For

, H0

dene the analytic function

f (z ) = 2z az , z .
Then Also

f is bounded in the strip S of the f (1/2 + it) = it 1/2 , so that

previous lemma and

f (0) = a,

f (1/2 + it) + f (1/2 + it) = 2it it Ja J it , .


So by the previous lemma we are done.

H and a cyclic and separating vector for M . Suppose S is the closure of x x on M . 1/ 2 Let = S S , and J be the antiunitary of the polar decomposition S = J .
Let

Theorem 13.2.3

be a von Neumann algebra on

Then (i) (ii)

JM J = M it M it = M t R a M
we know that

Proof. If

2it

. Conjugating by a unitary

it Ja J it dt M 2cosht
and writing

u M

= e2

we see that the

Fourier transforms of the strongly continuous rapidly decreasing functions it Ja J it it Ja J it and u u are equal. Hence it Ja J it M for all real

2cosht 2cosht t since it commutes with every unitary u M

. (Take inner products

with vectors if you are not happy with Fourier transforms of operator valued functions.) Putting

t=0

we see

JM J M

and by symmetry

JM J M

. Hence

JM J = M

and we are done.

89

Denition 13.2.4

The operator

of the previous result is called the mod-

ular conjugation and the strongly continuous one-parameter group of auto it it morphisms of M dened by t (x) = x is called the modular automorphism group dened by

13.3 Examples.
Example 13.3.1
ITPFI
The acronym ITPFI stands for innite tensor product of nite type I. These von Neumann algebras are formed by taking the *-algebra m union

A
in

as the

of tensor products

Am =

Mnk (C), the inclusion of Am

Am+1

being diagonal. The state each

k=1 on A is then the tensor product of states on

Mnk .

One may then perform the GNS construction with cyclic and

separating vector

given by

1 A ,

to obtain the von Neumann algebra

M =

Mnk (C)

as the weak closure of

acting on

H .

The case where

k=1 all the nk are equal to

and all the states are the same is called the Powers

factor and the product state the Powers state as it was R.Powers who rst showed that they give a continuum of non-isomorphic type III factors. A slight snag here is that we do not know that on

denes a faithful state

M.

But if we proceed anyway to construct what have to be

will soon see that the state is indeed faithful, i.e. Recall from exercise 13.0.6 that, for

is cyclic for

J and M .

we

Mn (C), and h (x) = trace(xh) where h is the diagonal matrix (density matrix) with hii = i , i = 1, i > 0, j i e and n (eij ) = eij (where dependence on h has been then Jn (eij ) = i ji j
suppressed). To diagonalise the modular operators on vincing to choose an orthonormal basis

completely it is most con-

di of the diagonal matrices, with d1 = 1. Then a basis for the Hilbert space H is formed by tensors k=1 vk where vk = 1 for large k , and is otherwise a di or an eij with i = j . We can guess that J is, on each basis vector, the tensor product of the J 's
coming from the matrix algebras. Dening it as such it is clearly an isometry

x A , JxJ is in M by the nite dimensional calculation! But the linear span of these JxJ is dense so is cyclic for M and hence separating for M . We are hence in
on

and thus extends to all of

H .

But then, for any

a position to apply Tomita Takesaki theory. Each of the basis elements is in M so S ( k=1 vk ) = k=1 wk where wk is vk if vk is diagonal, and eji if 90

vk = eij .
that

So

JS

is diagonal and hence essentially self-adjoint. We conclude

J (x) = Jm (x)
and

and

(x) = m (x)

for

x Am ,

=
k=1

hk

Example 13.3.2
Let

Group-measure-space construction.

be a discrete group acting on the nite measure space

(X, )

pre-

serving the class of the nite measure . The Hilbert space of the crossed L (X, ) is L2 (X, ) 2 () and as we saw in chapter 11 the vector 1 id is a cylic and separating vector for M = L (X, ) . product

is preserved by the the Radon Nikodym theorem L1 functions h so that (h (y )) = (x) where (y ) = yd. We know that, if x L (X, ) then S (u x) = x u 1 . In general X we will not be able to completely diagonalise but the same argument as in the previous example gives that the domain of is
Since the class of guarantees positive

{f : L2 (X, ) :

on which

|h (x)f (x)|2 d(x) < }


X

(f )( ) = h f ( ),
and

1/2 (Jf )( ) = h f ( ).
We can now nally answer the question as to which sums elements of M = L (X, ) .

x u

dene

Theorem 13.3.3
is such that

With notation as above, if the function

x L (X, )

x u , interpreted as a matrix of operators as in section 11.2, denes a bounded operator, then that operator is in M = L (X, ) .

x u commutes with Jxu J And for this it suces to check that the 2 commutation holds on functions of the form f for f L . This is just
Proof. for all By 13.2.3 it suces to show that and

x L (X, )

a routine computation.

91

Show that example 13.3.1 is in fact a special case of this group-measure-space example in which L (X, ) is provided by the tensor products of the diagonal elements and the group

Exercise 13.3.4

is a restricted innite

Cartesian product of cyclic groups, constructed from the non-diagonal Conclude by the method of 11.2.15 that ITPFI algbras are factors.

eij 's.

This example brings to light a signicant inadequacy of our treatment of Tomita-Takesaki theory. We would like to treat the case where the measure of the space is innite. Although of course we could choose an equivalent To do this we would have nite measure, this choice may not be natural. factor is to the trace on a type

to consider the theory of weights which are to states as the trace on a II

II1

factor. We need the same notion in order But a

to understand the origin of the term modular used above as coming from the modular function on a non-unimodular locally compact group. reader to Takesaki's book [3]. serious treatment of weights would take many pages so we simply refer the

Example 13.3.5
If

Hecke algebras la Bost-Connes.

is a nite group let

ug

and

vg

be the unitaries of the left and right

regular representations respectively. If H is a subgroup, the projection pH = 1 2 (G) onto functions that are right translation hH vh projects from |H | invariant under

H,

i.e. functions on the quotient space

G/H .

Thus the so-

called quasi-regular representation of

on

G/H

is a direct summand of

the left regular representation and we have from EP7 of chapter 4 that the 2 commutant of the action of G on (G/H ) is pH (G)pH where (G) is the algebra generated by the right regular representation (of course isomorphic to

C).

This commutant is spanned by the

pH vg pH

which, thought of as The subalgebra

functions on cosets of

G,

are multiples of the characteristic functions of the double space

the space of H H bi-invariant 2 functions and we see it is the commutant of G on (G/H ). It is known as the Hecke algebra for the pair where

HgH which form the double coset (G) spanned by these double cosets is (G, H )

H \G/H .

and has a considerable role to play

in the representation theory of nite groups. A famous example is the case

is the general linear group over a nite eld and

is the group of

upper triangular matrices. The coset space is then the so-called ag variety and the Hecke algebra in this case leads to a lot of beautiful mathemtatics. See Bourbaki []. Nothing could better indicate how dierently things work for innite discrete groups than how the Hecke algebra works. Far from being direct summands, the quasiregular representations can be totally dierent from the left

92

regular representations and can even generate type III factors! These Hecke algebras give nice examples where the modular operators can be calculated explicitly.

Denition 13.3.6
(a)

A subgroup

of the discrete group

is called almost

normal if either of the two equivalent conditions below is satised.

gHg 1 H
orbits of

is of nite index in

for all

g G. H
(i.e. the

(b) Each double coset of

is a nite union of left cosets of

on

G/H

are all nite).

G one may construct operators in the commutant G on 2 (G/H ) as follows: Given an element x of G/H let x be the characteristic function of x. 2 These functions form an orthonormal basis of (G/H ). Moreover each vector x is cyclic for the action of G hence separating for the commutant. If D is a double coset of H dene TD by the matrix
If is almost normal in of the quasiregular representation of

(TD )x,y =
Clearly

1 0 H

if

y 1 x = D ;

otherwise.

TD

is bounded since

is almost normal and it obviously com-

mutes with the action of

G.

From the denition we have

TD = TD1 .
It is also easy to check that

TD TE =
F

nF D,E TF

where the structure constants are dened by

nF D,E =

#(E/H ) 0

if

F ED; TD 's the HeckeHvN (G, H ). The

otherwise.

We will call the von Neumann algebra generated by the

H G and write it HvN (G, H ) by H is faithful and normal, and TD H , TD H = 0 unless D = D so that the TD 's are orthogonal. It is thus easy to calculate the operators for the modular theory on H (note that this 2 is not (G/H )). We guess as usual that J (TD ) = (constant)TD1 and by counting cosets in double cosets (or sizes of orbits of H on G/H ) we nd
von Neumann algebra of the pair
vector state

dened on

93

that the constant has to be is diagonal on the basis

(#(D/H ))1/2 (#(H \D))1/2 . Thus as TD of H so essentially self-adjoint and (TD ) = #(H \D) TD #(D/H )

before

JS

with the obvious domain. Thus

t (TD ) =

#(H \D) #(D/H )

it

TD .

Examples of almost normal subgroups are not hard to nd. The classical example of Hecke himself is the case where the case of the translations.

G = SL(2, Q) and H = SL(2, Z).

In this case the Hecke algebra is abelian. Bost and Connes in [4] examined

ax + b group over the rationals with the subgroup being integer They showed that HvN (G, H ) in this case is a type III factor

and made a connection with prime numbers.

13.4 The KMS condition.


In the examples of the previous section the operators of the modular group were easy to calculate explicitly, including the domain of

One can imagine

that this is not always so. If we are particularly interested in the modular group t it would be useful to be able to guess it and show that the guess is right without bothering about the domain of

The KMS (Kubo-Martin-

Schwinger) condition from quantum statistical mechanics allows us to do just that. The modular group came from the non-trace-like property of a state and the KMS condition allows us to correct for that. numbers (remember that Let us do a formal calculation assuming that the modular group can be extended to complex

is xed by

S, J

and

):

(xy ) = = = =
We conclude that

y , x y , J 1/2 x1 x1 , Sy y x1 , .

(xy ) = (yi (x)).


Thus the trace is commutative provide we operate by the modular group.

94

Exercise 13.4.1

If M is nite dimensional and is a faithful state, show that t = and that for each x and y in M there is an entire function F (z ) with, for t R,

F (t) = (t (x)y ) F (t + i) = (yt (x)).


If

and

is innite dimensional we would not expect the function

F (z )

of the

previous exercise to be entire.

Denition 13.4.2
on

Let

be a strongly continuous one parameter automor-

phism group of a von Neumann algebra

M , and be a faithful normal state satises the KMS condition for if t = and , for each x and y in M , there is a function F , continuous and bounded on the strip {z : 0 m(z ) 1}, analytic on the interior of the strip and such that for t R, M.
We say that
F (t) = (t (x)y ) F (t + i) = (yt (x)).

and

If is a faithful normal state on a von Neumann algebra M then t is the unique one parameter automorphism group satisfying the KMS condition for .
This chapter has been heavily technical so we defer the proof, which is by approximation on dense subspaces of the domain of an interesting corollary, identifying a part or trace.

Theorem 13.4.3

to which

the previous

calculations can be applied, to an appendix. We content ourselves here with

on which

behaves as a

Corollary 13.4.4
1. 2.

For

aM

the following are equivalent:

(ax) = (xa)
t (a) = a

for all

x M.

for all

t R.

2) Observe that for x M , x , a = , xa = , ax (by 1). So Sx, a = a , x so that a D (S ) and S (a) = . So (a) = a, it a = a and nally t (a) = a for all t R.
Proof.(1 (2

1) (t (x)a) = (t (xa)) = (xa) so that F (t) is constant.

Use the

Schwarz reection principle to create a holomorphic function, constant on in the union of the strip with its complex conjugate. Thus the strip and

R,

is constant on

(xa) = (ax).
95

Denition 13.4.5
.

The von Neumann subalgebra of

dened by either of

the conditions of the previous corollary is called the centraliser of the state

96

Chapter 14 Connes' theory of type III factors.


14.1 The Connes unitary cocycle Radon-Nikodym theorem.
This result will allow us to extract information from the modular group of a state which is independent of the state.

Theorem 14.1.1
algebra

Let

and

be faithful normal states on a von Neumann

M.

Then there is a strongly continous map

t ut

from

to the

unitary group of

so that
t = Adut t

t R.
ut t (us ) = ut+s .

Morevoer

ut

satises the cocycle condition

on M M2 (C) by ((x)ij ) = 1 0 1 ((x11 ) + (x22 )). The projection p = is xed by by 13.4.4. 2 0 0 So denes a one parameter automorphism group of pM M2 (C)p which satises the KMS condition for . Hence t (x e11 ) = t (x) e11 . Similarly 0 0 t (x e22 ) = t (x) e22 . Let Vt = t (1 e21 ). Then Vt Vt = and 0 1 1 0 0 0 Vt Vt = . Hence Vt = for some unitary vt M . Routine 0 0 vt 0
Proof. We dene the faithful normal state computations give the rest.

Corollary 14.1.2
M
is of type III.

If

is a factor and

is outer for any

phi

and

then

97

Proof. By the previous result it suces to exhibit a single faithful normal state on a type II factor with inner modular group. In the II1 case use the trace and in the II case choose a faithful normal state

on

B (H)

and use

tr ,

using the KMS condition (if necessary) to very that the modular

group for the tensor product is the tensor product of the modular groups.

Corollary 14.1.3 Denition 14.1.4


variant of

The subgroup of all

tR

for which

is inner is inde-

pendent of the faithful normal state

The subgroup of the previous corollary, which is an in-

M,

is called

T (M ). T (M ) for the Powers factor R where this refers all nk = 2 and all states having the same density

We shall now calculate to the ITPFI factor with 1 0 1+ . matrix h = 0 1+

Theorem 14.1.5
T (R ) =
.

2 Z log 2 = id
log

Proof. By the formula for the modular group

so

2 Z log

T (R ).
of the

For the other direction it suces to show that an automorphism form

= k=1 Adu
is outer whenever the unitary For this rst dene and observe that if = Adv then (uk 1 1) v = id on the matrix algebra Ak = k 1 M2 (C). By exercise 5.3.3 this means that v = uk w . Now it is clear from our basis that we can choose p j =1 xi 1 with non-zero inner procuct with v . But then xing p and letting k tend to innity we see that

u is uk = k 1u

not a scalar.

(p j =1 xi 1), v =
j =1
The left hand side does not depend on

xi , u 1, u k
and

k p

1, w .

| 1, u | = 1

which means that

is a scalar

| 1, w | 1 so we must have multiple of 1 by the Cauchy-

Schwarz inequality. We conclude that the Powers factors non-isomorphic for dierent values of

are type III factors ,mutually

98

14.2 Type III.


is easy to calculate for an ITPFI i as varies over j all the density matrices dening the product state. Apart from being closed factor. It is simply the closure of the set of all ratios under the inverse operation this set of non-negative real numbers has no particular structure and can be altered easily by making changes in nitely many density matrices which of course do not change the factor. The spectrum of the modular operator

Denition 14.2.1 Theorem 14.2.2 Theorem 14.2.3

If

is a von Neumann algebra the invariant

S (M )

is

the intersection over all faithful normal states responding modular operators A factor

of the spectra of their cor-

.
is of type III i

0 S (M ).
is a closed subgroup of

(Connes-van Daele)

S (M ) \ {0}

the positive real numbers.


There are only three kinds of closed subgroups of

R+ . 01
if

Denition 14.2.4

A factor

is called type III for

= 0 : S (M ) = {0} {1} 0 < < 1 : S (M ) = {0} {n : n Z} = 1 : S (M ) = {0} R+

Theorem 14.2.5

The Powers factor

is of type III .

In his thesis , Connes showed that every type III factor for action of If algebra whose generator scales the trace by

0<<1

is Connes thesis

canonically isomorphic to the crossed product of a type II factor with an

A is a locally compact abelian group with an action on a von Neumann on the crossed M , there is an action of the Pontryagin dual A product M A satisfying a (x) = x a (a)ua (ua ) = a
for if

xM ua are the

unitaries dening the crossed product.

The existence of the so-called dual action

is trivial proved since it is on L2 (A). implemented by the obvious unitary representation of A

99

A is nite consider the projection p = a ua M A. A Show that pM Ap = M p and thus show that (M A) A is isomorphic to M M|A| (C).
If
Observe that the crossed product of a von Neumann algebra M on H by the modular group does not depend, up to isomorphism, on the faithful normal state . This follows from theorem 14.1.1 by dening the unitary 2 on L (R, H) by

Exercise 14.2.6

V f (t) = ut f (t)
where of

is another faithful normal state with unitary one-cocycle

ut .

Conju-

gating the operators that generate

by

one obtains the generators

R.

Theorem 14.2.7 Denition 14.2.8


weights of

The crossed product of

by the modular group admits a

trace satisfying the properties of 9.1.9 The action of

on

Z (M

R) is called the ow of

M.
(Takesaki duality) The crossed product

Theorem 14.2.9

(M
is isomorphic to the tensor product
Thus if

R)

M B (H)

for

H = L2 (R).

is a factor the ow of weights is ergodic.

Theorem 14.2.10
III1 :

If

is a factor of type III the ow of weights is

The trivial ow on a one point set if

is III1 .

III :

0<
III0 :

2 The transitive ow on the circle with period if < 1.

is of type III ,

Ergodic non-transitive if

is of type III0 .

Moreover any ergodic non-transitive ow arises as the ow of weights for some type III0 factor.

100

Chapter 15 Hyperniteness
Denition 15.0.11
*-subalgebras of A von Neumann algebra

on a separable Hilbert space

is called hypernite if there is an increasing sequence

An

of nite dimensional

which generates

as a von Neumann algebra .

15.1 The hypernite type II1 factor R


The rst main result about hyperniteness was proved by Murray and von Neumann in []. We will use

to denote the hypernite II1 factor whose

uniqueness they proved.

Theorem 15.1.1
II1 factor.

Up to abstract isomorphism there is a unique hypernite

||x||2 = tr(x x)1/2 on M . It is not hard to show that a von Neumann subalgebra N of M is strongly dense in M i it is dense in ||||2 . Given a subalgebra A of M and a subset S M
Sketch of proof. One works with the norm one says

S A
if for each

xS

there is a

yA

with

||x y ||2 < . >0


there is a nite dimen-

The hyperniteness condition then implies:

For every nite subset sional *-subalgebra

of

S M M with

and every

S A.
101

The next step is to show that the A in the preceeding condition can be n n chosen to be the 2 2 matrices for some (possibly very large) n. This part uses what might be described as II1 factor technique. One begins with

and approximates all its minimal projections {ei } by ones whose traces are n numbers of the form k/2 . The matrix units of A can be changed a little bit

|| ||2 so that, together with matrix units conecting projections of trace 1/2n less than the ei , they generate a 2n 2n matrix algebra containing, up to , the matix units of A. Perturbation of the matrix units will involve results
in of the form:

u M satises ||(uu )2 uu ||2 < v M with ||v u||2 < F ( ) (for some nice function f with f (0) = 0).
If
or:

then there is a partial isometry

If p and q are projections with ||pq ||2 < pq = 0 and ||q q || < F ( ).

then there is a projection

with

or:

If (a) (b) (c)

fij

are almost

nn

matrix units, i.e.

||fij fji ||2 < ||fij fkl j,k fil ||2 < ||1 n
n i=1

f ii||2 <
units

then there are only on and

n n matrix F (0) = 0.

eij

with

||eij fij || < F ( )

where

depends

Such results are proved by a skilful use of the polar decomposition and spectral theorem. Thus one shows that in a hypernite type II1 factor one has: Property * :

For every nite subset

SM

and every

>0

there is a

2n 2n

matrix subalgebra of

with

S A.
One may now proceed to prove the theorem by choosing a sequence nk nk

|| ||2 -dense

xk

in

and inductively constructing an increasing sequence of

matrix algebras

Ak

with the property that

For each

k = 1, 2, 3, ...,

{x1 , x2 , ..., xk }

Ak . 1/k

102

The union of the the the

Ak 's is clearly dense in || ||2 . This is enough to prove theorem since the Ak 's can be used to give an isomorphism of M with type II1 factor M2 (C) constructed in section 7.2. Ak+1 eij
from

To construct add matrix units containing the close to the

Ak

one needs to arrange for the new algebra Now use property * to obtain a

to contain the one already constructed. So to the elements for

Ak+1 .
In

x1 , x2 , ..., xk+1 , B almost

xi

and the matrix units, with

small enough to control sums

over the matrix units

eij .

we know there are approximate matrix units

eij

so by a technical lemma, exact matrix units

fij

close to the

eij .
the

Now choose a unitary close to the identity which conjugates the

fij

to

eij

and use this unitary to conjugate

superalgebra is the identity.

Ak+1

and it contains the

B to a superalgebra of Ak . This xi up to epsilon since u is close to

This completes the sketch of the proof. The technique involved is considered standard in von Neumann algebras and the details can be found in .

dixmier
If

is the group of nitely supported permutations of a countably innite set then vN (S ) = M2 (C).
Proof. The subgroups of show that

Corollary 15.1.2

permuting an increasing sequence of nite sets

vN (S )

is hypernite.

It is surprising at rst sight that the type II1 factor from an ergodic measure-preserving transformation each (1) (2)

L (X, ) Z obtained n N and

is hypernite. This can

be shown by Rokhlin's tower theorem which asserts that, for each

>0

there is a measurable subset for .

AX

with and

T i (A) T j (A) =

1 i < j n,

i (X \ n i=0 T (A)) <

A can be combined, with a little perturbation to get the identity, to get a n n


The unitary of the crossed product and the characteristic function of matrix algebra. Careful application of the tower theorem will allow one to get any element of L (X, ), and u1 , in this matrix algebra up to some . This was rst proved by Henry Dye in who went on to prove that in fact all Dye groups of polynomial growth give hypernite II1 factors in this way. The ultimate result in this direction is the celebrated Injective factors theorem of Connes who showed that hyperniteness for a von Neumann algebra

u1

on

H is equivalent to injectivity

which means there is a projection

in the Banach space sense of norm one from

B (H)

onto

M.

This theorem,

whose proof is a great, great tour de force, has a raft of corollaries, many of

103

which were open questions. Let us just mention the fact that it follows easily that any subfactor of

R.
as

It also implies that

R which is innite dimensional is in fact isomorphic to vN (), as well as L (X, ) is hypernite as soon

is amenable.

15.2 The type III case.


The complete classication of injective(=hypernite) factors is a triumph

Connes actions Connes Injective factors

of 20th.

century mathematics.

Connes showed in that there is only one for each scaling factor

trace-scaling automorphism of

R B (H)

conjugacy. Together with this shows that for each a unique injective factor of type III .

with

= 1 up to 0 < < 1 there is

krieger injective

Using results of Krieger in , his thesis and , Connes showed that hypernite type III0 factors are classied by their ow of weights (up to conjugacy of ows, not orbit equivalence). This means that there is a rather large number of III0 factors but their classication is in the realm of ergodic theory rather than von Neumann algebras. The remaining case of injective type III1 factors was solved by Haagerup

ueIIIone

in . There is just one such factor and a hypernite factor is generically of type III1 .

104

Chapter 16 Central Sequences.


16.1 Generalities.
Denition 16.1.1
is said to be trivial property If norm bounded sequence

M is a type II1 factor , a central sequence in M is a (xn ) with limn ||[xn , a]||2 = 0. A central sequence if limn ||xn tr (xn )id||2 = 0. M is said to have

if there is a central

The collection of all central sequences is clearly a C -subalgebra of (N, M ). If is a free ultralter on N, the subalgebra I of (N, M ) consisting of sequences with limn ||xn ||2 = 0 is a 2-sided ideal of (N, M ). Note also that M is embedded in (N, M ) as constant sequences.

With notation as above, the ultraproduct of M along is by I . It is written M . The algebra of ( -)central sequences is the centraliser M = M M of M in (N, M ). the quotient of

Denition 16.1.2

(N, M )

By compactness, the trace on

denes a trace on

by

tr((xn )) = lim tr(xn )


n
and by denition it is faithful on

M . M
is complete

Exercise 16.1.3
in

||

Show that the unit ball (in the C norm) of ||2 so that M and M are von Neumann algebras.

16.2 Central sequences in R


All models for exhibit central sequences in abundance. The most obvious situation is that of M2 (C). Fixing x M2 (C) we can dene the sequence 105

xn = 1 1 1...x 1 1... with the x in the nth. slot in the tensor product. For large enough n, xn will commute with any element in the algebraic tensor product so by the obvious (in the II1 case!) inequality ||[xn , a]|| 2||xn || ||a||2 we see that (xn ) is central and clearly non-trivial if x is not a scalar. Just as
clearly the central sequence algebra is non-commutative as we only need to choose

and

that do not commute and construct the sequences

(xn )

and

(yn )

as above. In fact it is not hard to show that

is a factor.

Theorem 16.2.1
Proof. If

The central sequence algebra

is a type II1 factor .

(xn )

represents an element

X R ,

16.3 Type II1 factors without property .


Theorem 16.3.1
(a) Let the identity and three elements

be an icc group possessing a subset not containing , and such that

= {1} 1
1

(b), then for

and

are mutually disjoint.

x vN (), ||x tr(x)id||2 14 max{||[x, u ]||2 , ||[x, u ]||2 , ||[x, u ]||2 }.

Proof.Write

as

x u .

We will frequently use the formula

2 ||[x, u ]||2 2 = ||u1 xu x|| =

|x x1 |2 .

x tr(x)id it suces to prove the result if tr(x) = 0 ||x||2 = 1 so that for such an x we must show 1 14 max{||[x, u ]||2 , ||[x, u ]||2 , ||[x, u ]||2 }. We rst make a simple estimate. If is any subset of and then
By replacing

by

and we may suppose

|x |2

|x1 |2 | =

(|x | + |x1 |)||x | |x1 || (|x | + |x1 |)(|x x1 |)


2||x||2 (

|x x1 |2 )1/2

106

so that if

{, , }

we have

|

where

|x |2

|x1 |2 | 2

= max{||[x, u ]||2 , ||[x, u ]||2 , ||[x, u ]||2 }. ||x||2 = 1: |x |2 +


2

Let us now rst overestimate

1 2

|x1 |2

|x | + 2 .

Now underestimate it:

|x |2 +
2

|x 1 |2 +

|x 1 |2

|x | 4 . y
from the inequalities

y= 1 3y 4
Let as desired.

2 |x | and eliminate to obtain

1 2y + 2

and

1/14

It is easy to come up with groups having subsets as in the previous theorem. For instance if

G = F2 ,

free on generators

and

h,

let

be the set

of all reduced words ending in a non-zero power of g . Let = g , = h and = h1 . The same works for more than two generators. We conclude:

Theorem 16.3.2
.

The type II1 factor

vN (Fn ) for n does not have property

107

108

Chapter 17 Bimodules and property T

109

110

Chapter 18 Subfactors

111

112

Chapter 19 Free probability

113

114

Appendix A Normal implies ultraweakly continuous.

115

116

Appendix B Proof of the KMS condition.


a faithful normal state on a von Neumann algebra M . Then the modular group t is the unique one parameter automorphism group of

Theorem B.0.3
M

Let

be

which satises the KMS condition for

Proof. Perform the GNS construction with canonical cyclic and separating 1/2 vector and modular operators S = J . Recall that f () = for any it it function of with f (1) = 1. In particular (t (x) = ( x , so t preserves

.
(t (x)y ) = it y , x

Now let us check the rest of the KMS condition. We have

and

(yt (x)) = = = =
So let

yt (x), 1/ 2 J t (x ), J 1/2 y 1/2 y , 1/2 it x 1/2it y , 1/2 x

= y , = x and let pn be the spectral projection for for 1 the interval [1/n, n] so that pn tends strongly to 1 and are bounded on pn H . The functions Fn (z ) = iz pn ,
are then entire and

|Fn (t) (t (x)y )| = | it (1 pn ), | ||(1 pn ) || || || |Fn (t + i) (yt (x))| = | 1/2it (1 pn ), 1/2 | ||(1 pn )1/2 || ||1/2 ||.

117

Hence the we are done.

Fn are bounded and continuous on the strip {z : 0 < mz < 1}

and converge uniformly on its boundary. By the Phragmen-Lindelof theorem Now let us prove uniqueness of the modular group with the KMS condition. Let

be another continous one-parameter automorphism group satisfy-

ing KMS for

. t
preserves

The fact that

means we can dene a strongly continous

one-parameter unitary group t ut by ut x = t (x). By Stone's theorem it it is of the form t D for some non-singular positive self-adjoint operator

A.

The goal is to prove that

D = .

As a rst step we construct a dense

set of analytic vectors in operators of the form

by Fourier transform. Let

be the set of all

(t)t (x)dx f

for all

functions

of compact support on

R.

The integral converges

strongly so

f (log(D))x =

is in and

(t)Dit (x)dx f

A. Thus the spectral projections of D are in the strong closure of A A is dense. Moreover z Dz x is analytic for x A since x is in the spectral subspace of A for a bounded interval. Also A is invariant z under D by the functional calculus. To compare with dene, for x and y in A, the entire function F1 (z ) = Diz y , x .
Let for

be the function, analytic inside the strip and continuous and bounded

on it, guaranteed for

and

by the KMS condition. Then if we dene

G(z )

1 mz 1

by

G(z ) =
Since

F (z ) F1 (z ) F (z ) F1 (z )

if if

mz 0; mz 0.

G is analytic for 1 < mz < 1, hence equal to 0, and since both F and F1 are continous on the strip, (yt (x)) = F (t + i) = F1 (t + i) = D1it y , x . In particular putting t = 0 we get F
and

F1

agree on the real line

Dy , x

= (yx) = x , y = J 1/2 x , J 1/2 y = 1/2 y , 1/2 x


118

So

1/2 y D

is in the domain of

1/2

and

y = Dy .

z agree on A. But multiplication by the function e + 1 is a linear isomorphism of Cc so by functional calculus (D +1)A = A which is thus dense. Since D + 1 is invertible by spectral theory, any (, (D + 1) )
Thus and

in the graph of of the

essentially self-adjoint on

D + 1 can be approximated by (An , (D + 1)An ). Thus D is A, and both and D are self-adjoint extensions restriction of D to this domain. Thus D = .

119

120

Bibliography
[1] F.M. Goodman, P. de la Harpe, and V.F.R. Jones, Coxeter graphs and

towers of algebras, Springer-Verlag, 1989.


[2] M. Reed and B. Simon, Methods of Modern Mathematical Physics I:

Functional Analysis, Academic Press, 1972.


[3] M. Takesaki, Theory of Operator Algebras II. EMS series on Operator Algebras, 2002. [4] J.-B. Bost and A.Connes, Hecke algbebras, type III factors and phase

transitions with spontaneous symmetry breaking in number theory, Selecta Math. (New Series),1, (1995), 411-457.

121