Taken from C*-algebras and Operator Theory by Gerald Murphy.

A useful way of thinking of the theory of C*-algebras is as “non-commutative topology”. This is justified by the correspondence between abelian C*-algebras and locally compact Hausdorff spaces given by the Gelfand representation. The algebras studied in this chapter, von Neumann algebras, are a class of C*-algebras whose study can be thought of as “non-commutative measure theory”. The reason for the analogy in this case is that the abelian von Neumann algebras are (up to isomorphism) of the form L^\infty(\Omega,\mu), where (\Omega,\mu) is a measure space.

The theory of von Neumann algebras is a vast and very well-developed area of the theory of operator algebras. We shall be able only to cover some of the basics. The main results of this chapter are the von Neumann double commutant theorem and the Kaplansky density theorem.

4.1 The Double Commutant Theorem

There are a number of topologies on B(H) (H a Hilbert space), apart from the norm topology, that play a crucial role, and each has valuable properties that the others lack. The two most important are the strong (operator) topology and the weak (operator) topology. This section is concerned with the former (we shall introduce the weak topology in the next section). One of the reasons for the usefulness of the strong topology is the “order completeness” property asserted in Vigier’s theorem which is analogous to the order completeness property of the order completeness property of the real numbers \mathbb{R}.

Henceforth, we shall be using results concerning locally convex spaces.

Let H be a Hilbert space, and x\in H. Then the function

p_x:B(H)\rightarrow\mathbb{R}^+,\,T\mapsto \|Tx\|,

is a semi-norm on B(H). The locally convex topology on B(H) generated by the separating family \{p_x\}_{x\in H} is called the strong topology on B(H). Thus, a net \{u_\lambda\} converges strongly to an operator T on H iff Tx=\lim_\lambda T_\lambda(x) for all x\in H. It follows that the strong topology is weaker than the norm topology on B(H) (This is a theorem in topology. Suppose that \tau_1,\,\tau_2 are topologies on a set X. Suppose further that whenever \{x_\lambda\}\subset X is a net convergent with respect to the topology \tau_1, then \{x_\lambda\} is also convergent with respect to the topology \tau_2. Then \tau_2\subset \tau_1 — \tau_2 is weaker than \tau_1.).

With respect to the strong topology, B(H) is a topological vector space, so the operations of addition and scalar multiplication are strongly continuous. This is not the case in general for the multiplication and involution operations.

Example 4.1.1

Let H be an infinite-dimensional Hilbert space with an orthonormal basis \{e_n\}_{n\geq 1}. Set U_n=|e_1\rangle\langle e_n|. If x\in H, then U_n(x)=\langle e_n|x\rangle |e_1\rangle, so \lim_{n\rightarrow\infty}\|U_n(x)\|=\lim_{n\rightarrow \infty}|\langle x,e_n\rangle|=0. Thus, the sequence \{U_n\} is strongly convergent to to zero in B(H). Now U^\star_n=|e_n\rangle\langle e_1, so \|U^\star_n(x)\|=|\langle x,e_1\rangle|. Therefore \lim_{n\rightarrow \infty}\|U_n^\star(x)\|=1 for x=e_1, so the sequence \{U_n^\star\} does not converge strongly to zero. This shows that the operation T\mapsto T^* on B(H) is not strongly continuous, and therefore the strong and norm topologies on B(H) do not coincide.

Observe also that \|U_n\|=\|e_1\|\|e_n\|=1, so the sequence \{U_n\} is not convergent to zero, and therefore the norm \|\cdot\|:B(H)\rightarrow \mathbb{R}^+ is not strongly continuous.

The operation of multiplication B(H)\times B(H)\rightarrow B(H)(T,S)\mapsto TS is not strongly continuous either (Cf. Exercise 4.3).

The preceding example shows that the strong topology behaves badly in some respects, but it also has some very good qualities, as we shall prove in the next theorem.

Let H be an arbitrary Hilbert space and suppose that \{T_\lambda\}_{\lambda\in\Lambda} is an increasing net in B(H)_{\text{sa}} that converges strongly to T (so T also belongs to B(H)_{\text{sa}}). Then T=\sup_\lambda T_\lambda and \langle Tx,x\rangle=\sup_\lambda\langle T_\lambda x,x\rangle (for x\in H). The corresponding statement for decreasing nets in B(H)_{\text{sa}} is also true. Both of these observations follow from the fact that if a net \{T_\lambda\}_{\lambda\in\Lambda} converges strongly to an operator T, then \langle Tx,y\rangle=\lim_\lambda\langle T_\lambda x,y\rangle (x,y\in H).

Theorem 4.1.1 (Vigier)

Let \{T_\lambda\}_{\lambda\in\Lambda} be a net of hermitian operators on a Hilbert space H. Then \{T_\lambda\} is strongly convergent if it is increasing and bounded above, or if it is decreasing and bounded below.


We prove only the case where \{T_\lambda\} is increasing, since the decreasing case can be got from multiplying by minus 1.

Suppose then that \{T_\lambda\} is increasing and bounded above. By truncating the net (that is, by choosing an index \lambda_0\in\Lambda and considering the truncated net \{T_\lambda\}_{\lambda\geq\lambda_0}) we may suppose that \{T_\lambda\} is also bounded below, by S say. We may further assume that all T_\lambda are positive (by considering the net \{T_\lambda-S\} if necessary). There is a positive number M such that \|T_\lambda\|\leq M. It follows that the increasing net \{\langle T_\lambda x,x\rangle\} is bounded above by M\|x\|^2, so this net is convergent (by monotone convergence). Using the Polarisation Identity

\langle T_\lambda x,y\rangle-\frac{1}{4}\sum_{k=0}^3i^k\langle T_\lambda(x+i^ky),x+i^ky\rangle,

we see that \{\langle T_\lambda x,t\rangle\} is a convergent net for all x,y\in H. Letting \sigma(x,y) denote its limit, it is easy to check that the function

\sigma:H^2\rightarrow\mathbb{C}, \,(x,y)\mapsto \sigma(x,y),

is a sesquilinear form on H. Moreover,

|\sigma(x,y)|=\lim_\lambda |\langle T_\lambda x,y\rangle|\leq M\|x\|\|y\|,

so \sigma is bounded. Hence, there is an operator T on H such that \langle Tx,y\rangle=\sigma(x,y) for all x,y (Theorem 2.3.6).  Clearly \|T\|\leq M (same theorem), T is hermitian, and T_\lambda\leq T for all \lambda\in\Lambda. Also

\|Tx-T_\lambda x\|^2=\|(T-T_\lambda)^{1/2}(T-T_\lambda)^{1/2}x\|^2

\leq \|T-T_\lambda\|\|(T-T_\lambda)^{1/2}x\|^2

\leq 2M\langle(T-T_\lambda)x,x\rangle,

(the 2M term comes from a triangle inequality) and \langle (T-T_\lambda)x,x\rangle=0 so Tx=\lim_\lambda x. Thus, \{T_\lambda\} converges strongly to T \bullet

Remark 4.1.1

If \{P_\lambda\} is a net of projections on a Hilbert space strongly convergent to an operator P, then P is a projection. For P is self-adjoint and \langle Px,y\rangle=\lim_\lambda\langle P_\lambda x,y\rangle=\lim_\lambda\langle P_\lambda x,P_\lambda y\rangle=\langle Px,Py\rangle=\langle P^2x,y\rangle, so P=P^2.

Theorem 4.1.2

Suppose that \{P_\lambda\}_{\lambda\in\Lambda} is a net of projections on a Hilbert space H.

  1. If \{P_\lambda\} is increasing, then it is strongly convergent to the projection of H onto the closed vector subspace \overline{\bigcup_\lambda P_\lambda H}.
  2. If \{P_\lambda\} is decreasing, then it is strongly convergent to the projection of H onto \bigcap_\lambda P_\lambda H.


Left as an exercise \bullet

Just as for normed vector spaces, we say a family \{x_\lambda\}_{\lambda\in\Lambda} of elements of a locally convex space is summable to a point x is the net \{\sum_{\lambda\in F}x_\lambda\}_F (where F runs over all non-empty finite subsets of \Lambda) is convergent to x, and in this case we write x=\sum_{\lambda\in\Lambda} x_\lambda.

Theorem 4.1.3

Let \{P_\lambda\}_{\lambda\in\Lambda} be a family of projections on a Hilbert H space that are pairwise orthogonal (that is P_{\lambda_1} P_{\lambda_2}=0 if \lambda_1\neq\lambda_2.) Then \{P_\lambda\} is summable in the strong topology on B(H) to a projection P, such that

\|Px\|=\left(\sum_{\lambda\in\Lambda}\|P_\lambda x\|^2\right)^{1/2}     (x\in H).

If P=I, then the map

H\rightarrow\bigoplus_\lambda P_\lambda(H)x\mapsto \{P_\lambda x\},

is a unitary.


If F is a finite non-empty subset of $latex\Lambda$, then P_F=\sum_{\lambda\in F}P_\lambda is a projection. Therefore, \{P_F\}_F is an increasing net of projections, hence strongly convergent to a projection P; that is, the family \{P_\lambda\} is strongly summable to P. Moreover

\|Px\|^2=\lim_F\|P_Fx\|^2=\lim_F\sum_{\lambda\in F}\|P_\lambda x\|^2=\sum_{\lambda\in\Lambda}\|P_\lambda x\|^2.

The observation concerning the case when P=I is clear \bullet

If C is a subset of an algebra A, we define its commutant C^\prime to be the set of all elements of A that commute with all the elements of C. Observe that C^\prime is a subalgebra of A (straightforward). The double commutant C^{\prime\prime} of C is (C^\prime)^\prime. Similarly, C^{\prime\prime\prime}=(C^{\prime\prime})^\prime. Always C\subset C^{\prime\prime} and C^{\prime}\subset C^{\prime\prime\prime} (let c\in C. If a\in C^\prime then ac=ca. That is c\in C^\prime. Now let a\in C^\prime so that for all c\in C, we have ac=ca. Now for any b\in C^{\prime\prime}ab=ba so a commutes with everything in C^{\prime\prime}. That is a\in C^{\prime\prime\prime}. The text claims that in fact C^\prime=C^{\prime\prime\prime}. Now let a\in C'''. Now ab=ba, for all b\in C''. However C\subset C'' so therefore ab=ba for all b\in C. That is a\in C' also.). If A is a normed algebra, then C^\prime is closed (with more structure we could show this as follows. Define f_c:A\rightarrow A by a\mapsto ac-ca. Now f_c^{-1}(\{0\}) is closed as is \bigcap_{c\in C}f_c^{-1}(\{0\})=C^\prime.). If A is a *-algebra and C is self-adjoint, then C^\prime is a *-subalgebra of A (straightforward).

Lemma 4.1.4

Let H be a Hilbert space and A a *-subalgebra of B(H) containing I_H. Then A is strongly dense in A^{\prime\prime}.


Let T\in A^{\prime\prime}, x\in H, and K=\overline{Sx|S\in A}. Then K is a closed vector subspace of H which is invariant, and therefore reducing (K and K^\perp invariant), for all S\in A, since A is self-adjoint. Thus, if P is the projection of H onto K, then P\in A^\prime, so PT=TP. Hence Tx\in K, and therefore there is a sequence \{S_n\}\subset A such that Tx=\lim_{n\rightarrow\infty} S_nx.

For each positive integer n the map

\varphi:B(H)\rightarrow B(H^{(n)})S\mapsto (\delta_{ij}S),

is a unital *-homomorphism, so \varphi(A) is a subalgebra of B(H^{(n)}) containing I_{H^{(n)}}. Moreover, \varphi(T)\in(\varphi(A))^{\prime\prime}, for if R\in (\varphi(A))^\prime and S\in A then \varphi(S)R=R\varphi(S) \Rightarrow R_{ij}V=VR_{ij}. Hence R_{ij}\in A^\prime, so TR_{ij}=R_{ij}T. Therefore, \varphi(T)R=R\varphi(T). Suppose now that x=(x_1,\dots,x_n)\in H^{(n)}. Then by the first paragraph of this proof there is a sequence \{S_m\}_{m\geq 0} in A such that \varphi(T)x=\lim_{n\rightarrow\infty}\varphi(S_m)x. Hence, Tx_j=\lim_{n\rightarrow } for j=1,\dots,n.

We show that this implies that T is in the strong closure of A. If W is a strong neighbourhood of T, we must show that W\cap A is non-empty, and to do this we may suppose that W-T is a basic neighbourhood of 0. Therefore, there are elements x_1,\dots,x_n\in H and a positive number \varepsilon such that.

W-T=\{S\in B(H):\|Tx_j\|<\varepsilon;\,j=1,\dots,n\}.

Hence, there is a sequence \{S_m\}_{m\geq0} in A such that

Tx_j=\lim_{m\rightarrow\infty}S_mx_j   j=1,\dots,n.

Consequently, for some N the operator S_N\in W, so W\cap A\neq\emptyset \bullet

Let H be a Hilbert space. If A is a strongly closed *-subalgebra of B(H), we call A a von Neumann algebra on H. Since the strong topology is weaker than the norm topology, a strongly closed set is also norm closed. Hence, a von Neumann algebra is a C*-algebra.

Obviously, B(H) is a von Neumann algebra on H, as is \mathbb{C}I_H. If (H_\lambda)_{\lambda\in\Lambda} is a family of Hilbert spaces and A_\lambda is a von Neumann algebra on H_\lambda for each index \lambda, then it is an easy exercise  to show that the direct sum \oplus_{\lambda} A_\lambda is a von Neumann algebra on \oplus_\lambda H_\lambda.

If A is a *-algebra on a Hilbert space H, then its commutant A^\prime is a von Neumann algebra on H (Let \{T_\lambda\} be a net in A^\prime, strongly convergent to T, then T\in A^\prime. This is a consequence of the fact that A' is strongly closed.)

If A is a von Neumann algebra on H and P is a projection in A, then PAP is a von Neumann algebra on H (PAP=\{PTP:T\in A\}. We want to show that this set is strongly closed. Let \{PT_\lambda P\}_{\lambda\in\Lambda} be a net in A strongly convergent to S. We want to show that S=PTP for some T in A. Strong operator convergence implies that for any x\in H we have that \lim_\lambda \|PT_\lambda Px\|=\|Sx\|. We know that S\in A, as A is strongly closed. The rest follows through as multiplication is continuous). Also, M_n(A) is a von Neumann algebra on H^{(n)}.

If H is infinite-dimensional, then K(H) is not a von Neumann algebra. To see this, let E be an orthonormal basis for H, and for each finite non-empty subset of E let P_F=\sum_{e\in F}|e\rangle\langle e|. Then P_F is a finite rank projection and the net \{P_F\}_{F} converges strongly to I_H on H. If K(H) were a von Neumann algebra, this would imply that it contains I, and so \text{dim }H<|infty, contradicting our assumption on H.

A fundamental result concerning von Neumann algebras is the following, known as the Double Commutant Theorem. 

Theorem 4.1.5 (von Neumann’s Double Commutant Theorem)

Let A be a *-algebra on a Hilbert space H and suppose that I_H\in A. Then A is a von Neumann algebra on H iff A=A^{\prime\prime}.


Immediate from Lemma 4.1.4  \bullet

The intersection of a family of von Neumann algebras on a Hilbert space is also a von Neumann algebra. Thus, for any set C\subset B(H) there is a smallest von Neumann algebra A containing C. We call A the von Neumann algebra generated by C. If C is self-adjoint and contains I_H, then A=C^{\prime\prime}. If in addition C consists of commuting elements, then A is abelian (for in this case C\subset C^\prime\Rightarrow A=C^{\prime\prime}\subset C^\prime\Rightarrow A\subset A^\prime \checkmark). This implies that there are non-trivial examples of abelian von Neumann algebras. We give an explicit example.

Example 4.1.2

Let X be a compact Hausdorff space, and suppose that \mu is a finite positive regular Borel measure on X. We saw in Example 2.5.1 (https://jpmccarthymaths.wordpress.com/2011/01/19/c-algebras-and-operator-theory-2-5-the-spectral-theorem/) that the map

L^\infty(X,\mu)\rightarrow B(L^2(X,\mu))\varphi\mapsto M_\varphi,

is an isometric *-homomorphism. It’s range A is a C*-subalgebra of B(H). Denote by B the C*-subalgebra of all multiplication operators on L^2(X,\mu) with continuous symbol. The commutant of B is A. Hence, A is a von Neumann algebra on the Hilbert space L^2(X,\mu). Since A\subset A^\prime (because A is abelian) and A^\prime\subset B^\prime=A, therefore A=A^\prime. Consequently, A=B^{\prime\prime}, so B is strongly dense in A be Lemma 4.1.4.

Let K be a closed vector subspace of a Hilbert space H and let P be the projection of H onto K. If T\in B(H), let T_P=T_K be compression (restriction) of T to K. It is easy to verify that the map

PB(H)P\rightarrow B(K)T\mapsto T_K

is a *-isomorphism (\checkmark).

If A is a *-algebra on H and P is in A^\prime, set A_P=\{T_P:T\in A\}.

Lemma 4.1.6

Let A be a *-algebra on a Hilbert space H, and P a projection in A^\prime. Then PAP and A_P are *-algebras on H and P(H), respectively, and the map

PAP\rightarrow A_P , T\mapsto T_P,

is  a *-isomorphism. Moreover, if also P\in A^{\prime\prime}, then (A^\prime)_P=(A_P)^\prime.


We only show that P\in A^{\prime\prime}\Rightarrow P\in (A^\prime)_P=(A_P)^\prime — the rest is left to the exercises.

Suppose that T\in (A^\prime)_P and S\in A_p. Then there exist \tilde{T} and \tilde{S}\in A^\prime and A, respectively, such that T=\tilde{T}_P and S=\tilde{S}_P. Now we have

\tilde{S}=\left(\begin{array}{cc}S&S_1\\S_2&S_3\end{array}\right)\in A, and

\tilde{T}=\left(\begin{array}{cc}T&T_1\\T_2\\T_3\end{array}\right)\in A',

so that \tilde{S}=\tilde{T}=\tilde{T}\tilde{S} which implies that ST+S_1T_2=TS+T_1S_2. With respect to this same basis 


and we have that P\tilde{T}=\tilde{T}P; equivalently


which implies that T_1=0=T_2 so that TS=ST. Therefore T\in(A_P)', so (A')_P\subset (A_P)'.

Conversely, suppose now that T\in (A_P)^\prime, and write T=\tilde{T}_P for some \tilde{T}\in PB(H)P. Let 


Then if \tilde{S}\in A


So we have

\tilde{T}\tilde{S}=\left(\begin{array}{cc}TS &TS_1\&0\end{array}\right), and

\tilde{S}\tilde{T}=\left(\begin{array}{cc}ST &0\\ S_2T &0\end{array}\right).

But P\in A', so

P=\left(\begin{array}{cc}I&0\&0\end{array}\right) commutes with \tilde{S}.

This implies that S_1=0=S_2 so that \tilde{T}\tilde{S}=\tilde{S}\tilde{T}Consequently, \tilde{T}\in A^\prime, and therefore T\in (A^\prime)_P. This shows that the inclusion (A_P)^\prime\subset (A^\prime)_P holds \bullet

The reader should be aware that some authors define a von Neumann algebra on a Hilbert space H to be a *-algebra A on H such that A=A^{\prime\prime}. This automatically ensures that I_H\in A. However, proofs appear to run more smoothly if von Neumann algebras are defined as we have done. Moreover, we can frequently reduce to the case where A=A^{\prime\prime}, by the trick explained in Remark 4.1.2. Using our definition von Neumann algebras are still unitial, but the unit may not be the identity of the underlying Hilbert space.

Theorem 4.1.7

If A is a non-zero von Neumann algebra, then it is unital.


Suppose that A acts on a Hilbert space H, and let \{u_\lambda\}_{\lambda\in\Lambda} be an approximate unit for A. By Theorem 4.1.1, \{u_\lambda\} converges strongly to a self-adjoint operator, U say, and obviously U\in A, since A is strongly closed. If x\in H and T\in A, then UTx=\lim_\lambda u_\lambda Tx=Tx, so UT=T. Hence, U is a unit for A \bullet

Remark 4.1.2

Let A be a von Neumann algebra on a Hilbert space H and let U be the unit of A. Of course, U is a projection in A^\prime. The map

A\rightarrow A_UT\mapsto T_U,

is a *-isomorphism (by Lemma 4.1.6), and A_U is a von Neumann algebra on U(H) containing I_{U(H)}, so A_U=(A_U)^{\prime\prime}. This device will be frequently used to reduce to the case where the von Neumann algebra is it’s own double commutant (we don’t necessarily have A=A^{\prime\prime} if I_H\not\in A).

If T is an operator on a Hilbert space H, then its range projection [T] is the projection of H on \overline{T(H)}. We have [T]=[(TT^*)^{1/2}], since \overline{T(H)}^{\perp}=\text{ker}(T^*)=\text{ker}(TT^*)^{1/2}=\overline{(TT^*)^{1/2}(H)}^\perp ( First equality. Suppose x\in \overline{T(H)}^\perp. That is for all y\in H we have \langle x,Ty\rangle=0\Leftrightarrow \langle T^*x,y\rangle=0\Leftrightarrow x\in\text{ker }T^*.  The rest by the polar decomposition of T^*).

Theorem 4.1.9

If A is a von Neumann algebra, then it contains the range projections of all of its elements.


Let A act on H and let T\in A. Since (TT^*)^{1/2}\in A, to show that the range projection of T is in A, we may suppose that T\leq 0. Obviously we may also suppose that T\geq I. Let T_n=T^{1/2^n} for n\in\mathbb{N}. Then \{T_n\} is an increasing sequence of positive elements in the closed unit ball of A, so by Theorem 4.1.1, \{T_n\} is strongly convergent to a positive operator, P say. If x\in H, then

\|(P^2-T_n^2)x\|\leq \|(P^2-T_nP)x\|+\|(T_nP-T_n^2)x\|


Therefore, \{T_n^2\} converges to P^2 strongly. But T_n^2=T_{n-1} for all n>0 so P=P^2.

The sequence \{T_n\} is in C^*(T), so P(H)\subset \overline{T(H)}. The continuous functions

\sigma(T)\rightarrow\mathbb{R}t\mapsto t^{1+1/2^n},

form an increasing sequence and converge pointwise to the identity function t\mapsto t, so by Dini’s Theorem, they converge uniformly. Therefore, by the functional calculus, T=\lim_{n\rightarrow\infty} T^{1+1/2^n}; that is T=\lim_{n\rightarrow\infty} TT_n. Hence, T=TP=PT, so \overline{T(H)}\subset P(H). Therefore [T]=P\in A \bullet

Theorem 4.1.10

Let A be a von Neumann algebra on a Hilbert space H and T and element of A with polar decomposition T=U|T|. Then U\in A.


Let V be a unitary in the unital C*-algebra A^\prime. Then \tilde{V}=V^*UV is a partial isometry on H such that T=\tilde{V}|T| and \text{ker}(\tilde{V})=\text{ker }T (The projection onto the initial space of \tilde{V} is


=V^*\underbrace{U^*U}_{=\text{projection onto initial sp. of }U}V

But the initial space of U is \text{ran }|T|, and this projection lies in A, so 


that is the initial space of \tilde{V} is the initial space of U. Therefore \text{ker }\tilde{V}=\text{ker }U=\text{ker }T). It follows, therefore, from the uniqueness of the polar decomposition that \tilde{V}=U, so U and V commute (\checkmark). But A^\prime is the linear span of its isometries (Suppose that a C*-algebra A is unital and let a\in A_{\text{sa}}. Then \{U_a=a+i(I-a^2)^{1/2}:a\in A_\text{sa}\} is a set of unitaries that spans that space.), so U must commute with all elements of A^\prime, and therefore U\in A^{\prime\prime}=(A+\mathbb{C}1)^{\prime\prime}. By Lemma 4.1.4 there is a net \{U_\lambda\}_{\lambda\in\Lambda} and a net \{\alpha_\lambda\}_{\lambda\in\Lambda}  in \mathbb{C} such that the net \{U_\lambda+\alpha_\lambda 1\}_{\lambda\in\Lambda} converges strongly to U on H. If P=[|T|], then by Theorem 4.1.9 P\in A. Since (1-P)(H)=\overline{|T|(H)}^\perp=\text{ker }|T|=\text{ker }T=\text{ker }U (\text{ker }T=\text{ker U} by Theorem 2.3.4), we have U(1-P)=0; that is, U=UP. Therefore, U is the strong limit of the net \{U_\lambda P+\alpha_\lambda P 1\} which lies in A, so U\in A  \bullet

Theorem 4.1.11

Suppose that  A is a von Neumann algebra on a Hilbert space H.

  1. A is the closed linear span of its projections.
  2. If I_H\in A and T is a normal element of A, then E(X)\in A for every Borel set X of \sigma(T), where E is the spectral resolution of the identity for T.
  3. If I_H\in A and S\in B(H), then S\in A if and only if S commutes with all the projections in A^\prime.


We may suppose in all cases that I_H\in A (certainly in conclusions 2 and 3. I imagine we will be adapting Remark 4.1.2 to conclusion 1. if I_H\not\in A. ). We prove condition 2. first. Let T be a normal element with spectral resolution of the identity for T denoted by E. If S\in A^\prime, then ST=TS and ST^*=T^*S (as A^\prime is self-adjoint — take adjoints to S^*T=TS^*), so Sf(T)=f(T)S for every f\in B_\infty(\sigma(T)) (go on the Stone-Weierstrass!). In particular, S\,E(X)=E(X)\,S for every Borel set X of \sigma(T). Therefore, E(X)\in A^{\prime\prime}=A.

Condition 1. follows directly from condition 2., using the fact the closed linear span of the characteristic functions \{\chi_X:X\text{ a Borel set of }\sigma(T)\} is B_\infty (\sigma(T)) for each normal element T of A.

Condition 3. follows immediately from condition 1., since A^\prime is a von Neumann algebra, and therefore it is the closed linear span of its projections \bullet

We give an immediate and important application of this result  in the next theorem. First we make some observations.

Remark 4.1.3

If H is a Hilbert space, then B(H)^\prime=\mathbb{C}I_H. For it is obvious that \mathbb{C}^\prime=B(H), and since \mathbb{C} is a von Neumann algebra containing I_H, Theorem 4.1.5 implies that \mathbb{C}=\mathbb{C}^{\prime\prime}, so \mathbb{C}=B(H)^\prime.

Remark 4.1.4

If A is a C*-algebra acting on a Hilbert space H and S\subset H, consider the set AS:= \langle A(S)\rangle.

We say A acts non-degerately on H if \overline{AS}=H. Equivalently, for each non-zero x\in H there exists T\in A such that Tx\neq 0.

If A acts non-degenerately on H and \{U_\lambda\}_{\lambda\in\Lambda} is an approximate unit for A, then \{U_\lambda\} converges to I_H strongly. To show this we must prove that \lim_\lambda U_\lambda x=x, for all x\in H. This is clear if x=Ty for some T\in A and y\in H (because \overline{AH}=H — we can fill in the gaps by continuity). Now

\lim_\lambda U_n x=\lim_\lambda U_nTy=Ty=x \bullet

If A acts irreducibly on H and A\neq 0, then it acts non-degenerately, since \overline{AH} is a non-zero vector subspace  of H invariant for A, and therefore equals A (A acting irreducibly on H means that the only closed vector subspaces of H are \{0\} and H).

Theorem 4.1.12

Let A be a non-zero C*-algebra acting on a Hilbert space H. The following conditions are equivalent:

  1. A acts irreducibly on H.
  2. A^\prime=\mathbb{C}I_H.
  3. A is strongly dense in B(H).


If P is a projection in B(H), then P\in A^\prime if and only if the closed vector subspace P(H) of H is invariant for A. Since A^\prime is a von Neumann algebra, it is the closed linear span of its projections by Theorem 4.1.11, so if A acts irreducibly , then A^\prime has no projections except the trivial ones, and therefore A^\prime=\mathbb{C}I_H. Therefore 1. implies 2. The reverse implication is clear.

If we suppose now that A^\prime=\mathbb{C}I_H, then (A+\mathbb{C}I_H)^\prime=\mathbb{C}I_H (\checkmark), and therefore (A+\mathbb{C}I_H)^{\prime\prime}=B(H). By Lemma 4.1.4 the C*-algebra A+\mathbb{C}I_H is strongly dense in B(H). However, since A acts irreducibly on H, it acts non-degenerately, and therefore if \{U_\lambda\}_{\lambda\in\Lambda} is an approximate unit for A, it is strongly convergent to I_H. Therefore, I_H\in \overline{A}, and \overline{A}=\overline{A+\mathbb{C}I_H}=B(H), where the closure is strong closure (\checkmark). Hence 2. implies 3.

Finally, if A is strongly dense in B(H), then A^\prime=B(H)^\prime=\mathbb{C}, so 3. implies 2. \bullet

If P,Q are projections in a C*-algebra A, we say they are Murray-von Neumann equivalent, and we write P\sim Q, if there exists T\in A such that P=T^*T and Q=TT^*. It is indeed straightforward to show that this is indeed an equivalence relation on the set of projections of A (P=P^*P=PP^*=P so \sim is reflexive. Clearly \sim is symmetric. Now suppose P=T^*T and Q=TT^*; and Q=S^*S and R=SS^*. I don’t think I need to full result actually…). The relation \sim is of fundamental importance in the classification theory of von Neumann algebras (see the Addenda), and in K-theory for C*algebras. However, for the moment we only need one small result concerning \sim.

Remark 4.1.5

If P,Q are infinite-rank projections on a separable Hilbert space H, then P\sim Q (I can see how this goes back to the talk on E-semigroups.). To see this, choose orthonormal basis \{e_n\}_{n\geq 1} and \{f_n\}_{n\geq 1} for P(H) and Q(H). Let U:P(H)\rightarrow Q(H) be the unitary such that Ue_n=f_n for all n. Define T\in B(H) by setting T=U on P(H) and T=0 on (I_H-P)(H). It is easily verified that P=T^*T and Q=TT^*.

Theorem 4.1.15

If H is a separable infinite-dimensional Hilbert space, then K(H) is the unique non-trivial closed ideal of B(H).


Let I be a non-zero closed ideal of B(H). By Theorems 2.4.5 and 2.4.7, K(H)\subset I. Now if I\not\subset K(H), then by Corollary 4.1.14 (not covered as it uses hereditary C*-algebras), K(H) does not contain all of the projections of I. Hence, I has an infinite-rank projection, P say. If Q is any other infinite-rank projection on H, then as we say in Remark 4.1.5 there is an element T\in B(H) such that P=T^*T and Q=TT^*. Therefore, Q=Q^2=TPT^*, so Q also belongs to I. Hence, I contains all the projections of B(H), whether their rank is finite or infinite, and therefore I=B(H). Thus, we have shown that the only closed ideals of B(H) are \{0\}K(H) and B(H) \bullet

Remark 4.1.6

Let H be a separable infinite-dimensional Hilbert space. If \{x_n\}_{n\geq 1} is a dense sequence on H, it is easy to check that K(H) is the closed linear span of the operators |x_n\rangle\langle x_m| (m,n\geq 1) using the fact that K(H) is the closed linear span of the rank-one projections (Theorems 2.4.5 and 2.4.6). Hence K(H) is separable. However, B(H) is non-separable (and therefore B(H)/K(H) is non-separable). To see this, choose an orthonormal basis \{e_n\}_{n\geq 1} for H. For each set S of positive integers, let P_S be the projection onto \langle e_i:i\in S\rangle. Clearly, \|P_S-P_{S^\prime}\|=1 if S\neq S^\prime (\checkmark). Hence, the family of operators \{P_S\}_{S\subset 2^{\mathbb{N}}} cannot be in the closure of the range of any sequence in B(H).

Theorem 4.1.16

If H is an infinite-dimensional separable Hilbert space, then the Calkin algebra B(H)/ K(H) is a simple C*-algebra.


Let \mathcal{C} denote the Calkin algebra and \pi:B(H)\rightarrow\mathcal{C} the quotient map. To see that \mathcal{C} is simple, let I be a closed ideal in \mathcal{C}. Then \pi^{-1}(I) is a closed ideal in B(H) (as the quotient map is a homomorphism), and therefore by Theorem 4.1.15, \pi^{-1}(I)=\{0\},K(H) or B(H). Hence, I=\{0\} or \mathcal{C} \bullet

Remark 4.1.7

Let T,S be operators on a Hilbert space H such that TT^*\leq SS^*. Then there is an operator R\in B(H) such that T=SR. To see this, observe that

\|T^*x\|^2=\langle TT^*x,x\rangle\leq \langle SS^*x,x\rangle-\|S^*x\|^2.

Hence, we get a well-defined norm-decreasing linear map R_0:S^*(H)\rightarrow H by setting R_0S^*x=T^*x. Clearly, we can extend R_0 to a bounded linear R_1 on H. Setting R=R_1^*, we get the required result T=SR \bullet