This course studies von Neumann algebras as the operator-algebraic framework for functional analysis and quantum theory. The central object is a concrete algebra of bounded operators on a [Hilbert space](/page/Hilbert%20Space) that is stable under weak operator limits, and much of the course is devoted to understanding how such algebras are characterized, how they act on Hilbert spaces, and how their intrinsic structure is reflected in projections, partial isometries, states, and traces. Along the way, the course connects abstract algebraic ideas with geometric and analytic ones, showing how operator algebras encode symmetries, decompositions, and measurement-like phenomena. The chapters are arranged to build the theory in layers. The first part develops concrete operator algebras through weak operator closure, commutants, and the bicommutant theorem, then uses projections and partial isometries to expose the internal geometry of these algebras. Next, normal representations and preduals provide the duality theory needed for analysis, while Murray-von Neumann equivalence and comparison lead to the structure theory of factors and centers. The later chapters classify von Neumann algebras by type, distinguishing type I, II, and III behavior through minimal projections, traces, and the failure of trace, and then connect these structural results to quantum mechanics, conditional expectations, and subfactors before closing with a synthesis of the classification landscape and further directions. # Introduction These opening notes set the language and direction of the course rather than developing the full theory at once. The central objects are concrete algebras of bounded operators on a Hilbert space, but the topology is no longer the norm topology familiar from $C^*$-algebras. The guiding theme is that weak closure turns projections, commutants, and normal states into the main tools for analysis. Chapters 2, 5, and 6--9 make these ideas precise through the bicommutant theorem, the comparison theory of projections, and the type classification of factors. ## Why Weak Closure Enters Operator Algebra What changes when we replace norm approximation by convergence tested against vectors or functionals? In norm-closed operator algebras, approximation controls uniform behaviour on the whole unit ball of the Hilbert space. Von Neumann algebras are built for a different kind of limit: limits that are visible through matrix coefficients, spectral projections, and expectations of observables in states. The basic setting is a complex Hilbert space $H$ and the algebra $\mathcal{L}(H)$ of bounded linear operators on $H$. The weak operator topology is designed to remember all scalar quantities of the form $(Tx,y)_H$, where $x,y \in H$. This is weaker than norm convergence, but it is strong enough to retain the projection structure that spectral theory produces. [definition: Weak Operator Topology] Let $H$ be a complex Hilbert space. The weak operator topology on $\mathcal{L}(H)$ is the weakest topology for which, for every $x,y \in H$, the map $T \mapsto (Tx,y)_H$ from $\mathcal{L}(H)$ to $\mathbb C$ is continuous. [/definition] This definition says that an operator net converges weakly exactly when every matrix coefficient converges. The use of nets rather than sequences is part of the subject: weak operator closures are topological closures in a topology that is often not first countable on large bounded sets. [example: Diagonal Operators] Let $H=\ell^2(\mathbb N)$ with standard [orthonormal basis](/page/Orthonormal%20Basis) $(e_n)_{n=1}^{\infty}$. For $a=(a_n)_{n=1}^{\infty} \in \ell^\infty(\mathbb N)$, define $D_a e_n=a_n e_n$ and extend linearly and continuously to $H$. If $x=\sum_{n=1}^{\infty}x_n e_n$, then \begin{align*} D_a x=\sum_{n=1}^{\infty}a_n x_n e_n. \end{align*} Thus $D_aD_b e_n=D_a(b_n e_n)=b_n a_n e_n$, so $D_aD_b=D_{ab}$, and $D_aD_b=D_bD_a$ because $a_nb_n=b_na_n$ for every $n$. Also $D_a^*e_n=\overline{a_n}e_n$, since $(D_a e_n,e_m)_H=a_n\delta_{nm}=(e_n,\overline{a_m}e_m)_H$. Hence the diagonal operators form a commutative $*$-algebra, and the map $a \mapsto D_a$ identifies it with $\ell^\infty(\mathbb N)$. Now let $(D_{a^\lambda})_{\lambda \in \Lambda}$ be a bounded net of diagonal operators, with $\sup_\lambda \|D_{a^\lambda}\| \le C$. If $D_{a^\lambda}$ converges weakly to $D_a$, then for each $n$, \begin{align*} a^\lambda_n=(D_{a^\lambda}e_n,e_n)_H \longrightarrow (D_a e_n,e_n)_H=a_n. \end{align*} So weak operator convergence implies coordinatewise convergence of the diagonal entries. Conversely, suppose $a^\lambda_n \to a_n$ for every $n$ and $\sup_\lambda\|D_{a^\lambda}\|\le C$. Then $|a_n|\le C$ for every $n$, so $a\in \ell^\infty(\mathbb N)$. For $x=\sum x_ne_n$ and $y=\sum y_ne_n$ in $\ell^2(\mathbb N)$, \begin{align*} (D_{a^\lambda}x,y)_H-(D_ax,y)_H=\sum_{n=1}^{\infty}(a^\lambda_n-a_n)x_n\overline{y_n}. \end{align*} Choose $N$ so that \begin{align*} 2C\left(\sum_{n>N}|x_n|^2\right)^{1/2}\left(\sum_{n>N}|y_n|^2\right)^{1/2}<\varepsilon/2. \end{align*} For the finite part, coordinatewise convergence gives an index $\lambda_0$ such that whenever $\lambda\ge \lambda_0$, \begin{align*} \left|\sum_{n=1}^{N}(a^\lambda_n-a_n)x_n\overline{y_n}\right|<\varepsilon/2. \end{align*} For the tail, Cauchy-Schwarz gives \begin{align*} \left|\sum_{n>N}(a^\lambda_n-a_n)x_n\overline{y_n}\right|\le 2C\left(\sum_{n>N}|x_n|^2\right)^{1/2}\left(\sum_{n>N}|y_n|^2\right)^{1/2}<\varepsilon/2. \end{align*} Therefore $(D_{a^\lambda}x,y)_H \to (D_ax,y)_H$ for all $x,y\in H$, which is exactly weak operator convergence. Thus, on bounded diagonal nets, the weak operator topology is coordinatewise convergence of the diagonal entries, recording pointwise spectral data rather than [uniform convergence](/page/Uniform%20Convergence) of the sequences. [/example] This example is the first model to keep in mind. Norm closure would ask for uniform approximation of the diagonal sequence, while weak closure permits limits detected coordinate by coordinate. The same phenomenon later becomes $L^\infty(X,\mu)$ acting by multiplication on $L^2(X,\mu)$. ## Concrete Algebras and Their Adjoints Which operator algebras are stable enough to carry spectral projections and adjoints? A general subalgebra of $\mathcal{L}(H)$ may fail to contain the projections generated by the spectral theorem, so it does not interact well with the geometry of $H$. The algebras in this course are unital, closed under adjoints, and closed in an operator topology adapted to Hilbert-space matrix coefficients. [definition: Von Neumann Algebra] Let $H$ be a complex Hilbert space. A von Neumann algebra on $H$ is a unital $*$-subalgebra $M \subseteq \mathcal{L}(H)$ that is closed in the weak operator topology. [/definition] The definition is concrete: the Hilbert space representation is part of the data. Abstract $C^*$-algebras appear in the background, but this course studies the extra structure that emerges after representing an algebra on $H$ and closing it weakly. [example: Full Operator Algebra] Let $H$ be a complex Hilbert space. The set $\mathcal{L}(H)$ is an algebra under the usual operator operations: if $S,T \in \mathcal{L}(H)$ and $\alpha,\beta \in \mathbb C$, then $\alpha S+\beta T \in \mathcal{L}(H)$, and $ST \in \mathcal{L}(H)$ because \begin{align*} \|STx\| \le \|S\|\|Tx\| \le \|S\|\|T\|\|x\| \end{align*} for every $x \in H$. The identity operator $I_H$ belongs to $\mathcal{L}(H)$, so the algebra is unital. If $T \in \mathcal{L}(H)$, then its Hilbert-space adjoint $T^*$ is also bounded and satisfies \begin{align*} \|T^*\|=\|T\|. \end{align*} Thus $\mathcal{L}(H)$ is a unital $*$-subalgebra of itself. It remains only to check weak operator closedness. A net $(T_\lambda)$ in $\mathcal{L}(H)$ converges in the weak operator topology precisely when there is some operator $T \in \mathcal{L}(H)$ such that, for all $x,y \in H$, \begin{align*} (T_\lambda x,y)_H \to (Tx,y)_H. \end{align*} Since the limiting operator in this topology is, by definition, an element of the ambient space $\mathcal{L}(H)$, every weak operator limit of a net from $\mathcal{L}(H)$ still lies in $\mathcal{L}(H)$. Therefore $\mathcal{L}(H)$ is weak operator closed, hence a von Neumann algebra on $H$. It is the largest possible von Neumann algebra on the fixed Hilbert space $H$, and later it appears as the basic type $I$ model. [/example] The opposite extreme is just as important: commutative von Neumann algebras behave like algebras of [measurable functions](/page/Measurable%20Functions). This parallel is one reason measure theory is a prerequisite rather than an optional background topic. [example: Multiplication Algebra] Let $(X,\mathcal E,\mu)$ be a [measure space](/page/Measure%20Space) and set $H=L^2(X,\mathcal E,\mu)$. For $f\in L^\infty(X,\mathcal E,\mu)$ and $g\in H$, the product $fg$ belongs to $L^2$ because \begin{align*} \|fg\|_2^2=\int_X |f|^2|g|^2\,d\mu \le \|f\|_\infty^2\int_X |g|^2\,d\mu=\|f\|_\infty^2\|g\|_2^2. \end{align*} Thus $M_f g=fg$ defines a bounded operator on $H$. If $f,h\in L^\infty$ and $g\in H$, then \begin{align*} M_fM_hg=M_f(hg)=f(hg)=(fh)g=M_{fh}g. \end{align*} Since $fh=hf$ almost everywhere, $M_fM_h=M_hM_f$. Also $M_1=I_H$, and for $g,k\in H$, \begin{align*} (M_f g,k)_H=\int_X f g\overline{k}\,d\mu=\int_X g\overline{\overline f k}\,d\mu=(g,M_{\overline f}k)_H. \end{align*} Therefore $M_f^*=M_{\overline f}$, so $\{M_f:f\in L^\infty(X,\mathcal E,\mu)\}$ is a unital commutative $*$-subalgebra of $\mathcal L(H)$. It remains to see that it is weak operator closed. Suppose $M_{f_\lambda}$ converges weakly to some $S\in\mathcal L(H)$. By the uniform [boundedness theorem](/theorems/181), there is $C>0$ such that $\|M_{f_\lambda}\|\le C$ for all $\lambda$. For each $u\in L^1(X,\mathcal E,\mu)$, write \begin{align*} u=|u|^{1/2}\overline{\overline{\operatorname{sgn}(u)}|u|^{1/2}}, \end{align*} where $\operatorname{sgn}(u)=0$ on $\{u=0\}$ and $\operatorname{sgn}(u)=u/|u|$ on $\{u\ne 0\}$. Both factors lie in $L^2$, and their $L^2$ norms multiply to $\|u\|_1$. Hence the limit \begin{align*} \Phi(u)=\lim_\lambda \int_X f_\lambda u\,d\mu \end{align*} exists, because it is a weak matrix coefficient of the net $M_{f_\lambda}$. Moreover \begin{align*} |\Phi(u)|\le C\|u\|_1. \end{align*} By $L^1$-$L^\infty$ duality, there is some $f\in L^\infty(X,\mathcal E,\mu)$ such that \begin{align*} \Phi(u)=\int_X fu\,d\mu \end{align*} for every $u\in L^1$. Taking $u=g\overline{k}$ with $g,k\in L^2$, we get \begin{align*} (Sg,k)_H=\lim_\lambda (M_{f_\lambda}g,k)_H=\lim_\lambda\int_X f_\lambda g\overline{k}\,d\mu=\int_X fg\overline{k}\,d\mu=(M_f g,k)_H. \end{align*} Since this holds for every $k\in H$, $Sg=M_fg$ for every $g\in H$, so $S=M_f$. Thus the multiplication operators form a weakly closed unital commutative $*$-subalgebra, hence a commutative von Neumann algebra. Finally, its projections are exactly the multiplication operators by characteristic functions. If $M_f$ is a projection, then $M_f^*=M_f$ and $M_f^2=M_f$, so $f=\overline f$ and $f^2=f$ almost everywhere. A real number satisfying $t^2=t$ is either $0$ or $1$, hence $f=\mathbf 1_A$ almost everywhere for $A=\{x:f(x)=1\}$. Conversely, \begin{align*} M_{\mathbf 1_A}^2=M_{\mathbf 1_A\mathbf 1_A}=M_{\mathbf 1_A}, \qquad M_{\mathbf 1_A}^*=M_{\overline{\mathbf 1_A}}=M_{\mathbf 1_A}. \end{align*} So measurable sets give precisely the projections in this algebra, making its projection theory measure theory in operator form. [/example] The multiplication example motivates the later slogan that noncommutative von Neumann algebras behave like algebras of essentially bounded functions on noncommutative measure spaces. The course makes this slogan precise through projections, states, and traces rather than through points. ## Commutants as a Closure Mechanism How can weak closure be recognised without testing every possible weak limit? The key operation is to reverse the question: instead of asking which limits lie in an algebra, ask which operators commute with the algebra. Taking commutants turns invariant algebraic information into a closed operator algebra. [definition: Commutant] Let $S \subseteq \mathcal{L}(H)$. The commutant of $S$ is \begin{align*} S' := \{T \in \mathcal{L}(H) : TA=AT \text{ for all } A \in S\}. \end{align*} The double commutant of $S$ is $S'' := (S')'$. [/definition] The commutant is automatically weakly closed because each relation $TA=AT$ is detected by matrix coefficients. It is also a unital algebra closed under adjoints when $S$ is self-adjoint. This makes the next question unavoidable: does every weakly closed unital $*$-algebra arise from its commutation relations alone? [quotetheorem:1246] [citeproof:1246] This theorem is the structural reason the subject is tractable. It replaces a topological closure condition by an algebraic condition involving commutation, and it explains why projections onto invariant or reducing subspaces appear throughout the course. The hypotheses are part of the content: unitality and $*$-closure ensure that the algebra has enough reducing-subspace structure for commutants to detect its weak closure. For example, the non-self-adjoint algebra of upper triangular matrices in $M_2(\mathbb C)$ is norm closed, but its double commutant is not obtained by simply taking a weak closure of a von Neumann algebra, because it is not closed under adjoints and so cannot be a von Neumann algebra. The theorem also does not say that every weakly closed algebra is a von Neumann algebra; without the $*$-algebra hypotheses, weak closure alone does not supply the spectral projections and adjoints needed later. This limitation is why the course returns repeatedly to self-adjointness, projections, and reducing subspaces when using commutants in examples. [example: Matrix Algebra Acting with Multiplicity] Let $(e_1,\ldots,e_n)$ be the standard basis of $\mathbb C^n$, and let $E_{ij}\in M_n(\mathbb C)$ be the matrix unit satisfying $E_{ij}e_b=\delta_{jb}e_i$. Write any $T\in\mathcal L(\mathbb C^n\otimes K)$ in block form \begin{align*} T(e_b\otimes \eta)=\sum_{a=1}^n e_a\otimes T_{ab}\eta \end{align*} with operators $T_{ab}\in\mathcal L(K)$. Suppose first that $T$ commutes with every $E_{ij}\otimes I_K$. For $b\in\{1,\ldots,n\}$ and $\eta\in K$, \begin{align*} (E_{ij}\otimes I_K)T(e_b\otimes \eta)=e_i\otimes T_{jb}\eta. \end{align*} On the other hand, \begin{align*} T(E_{ij}\otimes I_K)(e_b\otimes \eta)=\delta_{jb}\sum_{a=1}^n e_a\otimes T_{ai}\eta. \end{align*} If $b\ne j$, equality gives $e_i\otimes T_{jb}\eta=0$ for every $i$ and $\eta$, so $T_{jb}=0$. Since $j$ and $b$ are arbitrary with $j\ne b$, all off-diagonal blocks vanish. If $b=j$, equality gives \begin{align*} e_i\otimes T_{jj}\eta=\sum_{a=1}^n e_a\otimes T_{ai}\eta. \end{align*} Comparing the $e_i$-coordinate gives $T_{jj}\eta=T_{ii}\eta$ for every $\eta$, so all diagonal blocks are one operator $B\in\mathcal L(K)$. Thus \begin{align*} T(e_b\otimes \eta)=e_b\otimes B\eta=(I_n\otimes B)(e_b\otimes \eta), \end{align*} and hence $T=I_n\otimes B$. Conversely, for $B\in\mathcal L(K)$, \begin{align*} (E_{ij}\otimes I_K)(I_n\otimes B)(e_b\otimes\eta)=\delta_{jb}e_i\otimes B\eta. \end{align*} Also, \begin{align*} (I_n\otimes B)(E_{ij}\otimes I_K)(e_b\otimes\eta)=\delta_{jb}e_i\otimes B\eta. \end{align*} So $I_n\otimes B$ commutes with every $E_{ij}\otimes I_K$, and therefore \begin{align*} A'=I_n\otimes \mathcal L(K). \end{align*} Now take $S\in A''$, so $S$ commutes with every $I_n\otimes R$, $R\in\mathcal L(K)$. Writing \begin{align*} S(e_b\otimes \eta)=\sum_{a=1}^n e_a\otimes S_{ab}\eta, \end{align*} the equality $S(I_n\otimes R)=(I_n\otimes R)S$ gives \begin{align*} \sum_{a=1}^n e_a\otimes S_{ab}R\eta=\sum_{a=1}^n e_a\otimes RS_{ab}\eta. \end{align*} Thus each block satisfies $S_{ab}R=RS_{ab}$ for every $R\in\mathcal L(K)$. An operator on $K$ commuting with all of $\mathcal L(K)$ must be scalar: if $0\ne \xi\in K$, then every rank-one operator $\eta\mapsto(\eta,\xi)_K\zeta$ commutes with it, forcing all vectors to be eigenvectors with the same eigenvalue. Hence $S_{ab}=\alpha_{ab}I_K$ for scalars $\alpha_{ab}\in\mathbb C$, and \begin{align*} S=\left(\sum_{a,b=1}^n \alpha_{ab}E_{ab}\right)\otimes I_K. \end{align*} Therefore \begin{align*} A''=M_n(\mathbb C)\otimes I_K=A. \end{align*} The commutant records exactly the unused multiplicity space $K$, and taking the commutant again recovers the original matrix algebra acting on the first tensor factor. [/example] The finite-dimensional computation already foreshadows the classification programme: commutants measure how much multiplicity remains after an algebra acts. Factors are the case where the centre contains no internal decomposition data. ## Projections, Equivalence, and Type What replaces dimension when projections live inside a noncommutative algebra? In finite-dimensional linear algebra, projections are compared by rank. In a von Neumann algebra, projections are compared by partial isometries belonging to the algebra itself, and this internal comparison leads to Murray-von Neumann equivalence. [definition: Projection] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra. A projection in $M$ is an operator $p \in M$ such that \begin{align*} p^2=p, \qquad p^*=p. \end{align*} [/definition] A projection is the operator-algebraic form of a closed subspace, but the condition $p \in M$ says that the subspace is visible to the algebra. To compare two visible subspaces, rank is not flexible enough in infinite algebras. The comparison must be made by an operator in $M$ that carries one projection range onto the other. [definition: Murray Von Neumann Equivalence] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra, and let $p,q \in M$ be projections. The projections $p$ and $q$ are Murray-von Neumann equivalent in $M$, written $p \sim q$, if there exists $v \in M$ such that \begin{align*} v^*v=p, \qquad vv^*=q. \end{align*} [/definition] The operator $v$ is a partial isometry inside the algebra, so the comparison is intrinsic to $M$ rather than to the whole Hilbert space. This distinction becomes essential for factors, where the same Hilbert-space projection can behave differently depending on the ambient von Neumann algebra. [example: Rank Comparison in $M_n(\mathbb C)$] In $M_n(\mathbb C)$ acting on $\mathbb C^n$, two projections are Murray-von Neumann equivalent exactly when they have the same rank. First suppose $\operatorname{rank}(p)=\operatorname{rank}(q)$. Since $p=p^*=p^2$ and $q=q^*=q^2$, they are orthogonal projections onto $p\mathbb C^n$ and $q\mathbb C^n$. Choose a unitary [linear map](/page/Linear%20Map) $U:p\mathbb C^n\to q\mathbb C^n$. Every $x\in\mathbb C^n$ has a unique decomposition $x=x_1+x_0$ with $x_1\in p\mathbb C^n$ and $x_0\in (p\mathbb C^n)^\perp$, and define \begin{align*} vx=Ux_1. \end{align*} Then $v^*$ is given by $v^*y=U^{-1}y_1$ when $y=y_1+y_0$ with $y_1\in q\mathbb C^n$ and $y_0\in(q\mathbb C^n)^\perp$. Hence, for $x=x_1+x_0$ as above, \begin{align*} v^*vx=v^*Ux_1=x_1=px. \end{align*} Similarly, for $y=y_1+y_0$ with $y_1\in q\mathbb C^n$ and $y_0\in(q\mathbb C^n)^\perp$, \begin{align*} vv^*y=vU^{-1}y_1=y_1=qy. \end{align*} Thus $v^*v=p$ and $vv^*=q$, so $p\sim q$. Conversely, suppose $p\sim q$, so there is $v\in M_n(\mathbb C)$ with $v^*v=p$ and $vv^*=q$. If $x\in p\mathbb C^n$, then $px=x$, and \begin{align*} \|vx\|^2=(vx,vx)=(v^*vx,x)=(px,x)=(x,x)=\|x\|^2. \end{align*} Also \begin{align*} q(vx)=vv^*vx=v(v^*vx)=v(px)=vx, \end{align*} so $v$ maps $p\mathbb C^n$ isometrically into $q\mathbb C^n$. To see that this map is onto, let $y\in q\mathbb C^n$. Then $qy=y$, so \begin{align*} y=qy=vv^*y=v(v^*y). \end{align*} Moreover, \begin{align*} p(v^*y)=v^*v v^*y=v^*q y=v^*y, \end{align*} so $v^*y\in p\mathbb C^n$. Hence every $y\in q\mathbb C^n$ lies in the image of $v|_{p\mathbb C^n}$. Therefore $p\mathbb C^n$ and $q\mathbb C^n$ have the same dimension, which means $\operatorname{rank}(p)=\operatorname{rank}(q)$. Thus Murray-von Neumann equivalence in finite matrices recovers ordinary rank comparison. [/example] Finite matrices therefore recover the familiar dimension theory. Infinite von Neumann algebras allow projections equivalent to proper subprojections, and that phenomenon is where the type $I$, $II$, and $III$ classification begins. ## States, Normality, and Operational Convergence Which linear functionals should count as physically or analytically meaningful states on a von Neumann algebra? In a $C^*$-algebra, states are positive norm-one functionals. For von Neumann algebras, the [weak topology](/page/Weak%20Topology) singles out a smaller and better-behaved class: normal states, which respect increasing limits of projections and operators. [definition: State] Let $M$ be a unital $C^*$-algebra. A state on $M$ is a linear functional $\varphi:M \to \mathbb C$ such that $\varphi(a^*a) \ge 0$ for all $a \in M$ and $\varphi(1_M)=1$. [/definition] States let us evaluate observables as expected values. In the concrete von Neumann algebra setting, vector states are the first examples and connect the abstract definition to Hilbert-space geometry. [example: Vector State] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra and let $\xi \in H$ satisfy $\|\xi\|=1$. Define \begin{align*} \omega_\xi(T)=(T\xi,\xi)_H, \qquad T \in M. \end{align*} If $S,T\in M$ and $\alpha,\beta\in\mathbb C$, then \begin{align*} \omega_\xi(\alpha S+\beta T)=((\alpha S+\beta T)\xi,\xi)_H=(\alpha S\xi+\beta T\xi,\xi)_H=\alpha(S\xi,\xi)_H+\beta(T\xi,\xi)_H=\alpha\omega_\xi(S)+\beta\omega_\xi(T). \end{align*} Thus $\omega_\xi$ is linear. For positivity, take $T\in M$. Since $M$ is a $*$-algebra, $T^*T\in M$, and \begin{align*} \omega_\xi(T^*T)=(T^*T\xi,\xi)_H=(T\xi,T\xi)_H=\|T\xi\|^2\ge 0. \end{align*} For the unit, \begin{align*} \omega_\xi(I_H)=(I_H\xi,\xi)_H=(\xi,\xi)_H=\|\xi\|^2=1. \end{align*} Therefore $\omega_\xi$ is a positive linear functional of value $1$ on the identity, hence a state on $M$. [/example] Vector states are continuous for the weak operator topology, but arbitrary states on the underlying $C^*$-algebra need not respect the monotone limits that von Neumann algebras are built to contain. This creates a mismatch between the order-completeness of a von Neumann algebra and the functionals used to test it: a positive functional should send bounded increasing suprema to ordinary suprema if it is to see the same limiting process as the algebra. The following definition names exactly that compatibility condition. [definition: Normal Positive Functional] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra. A positive linear functional $\varphi:M \to \mathbb C$ is normal if, whenever $(a_i)_{i \in I}$ is an increasing bounded net of self-adjoint elements of $M$ with least upper bound $a \in M$, we have \begin{align*} \varphi(a)=\sup_{i \in I} \varphi(a_i). \end{align*} [/definition] This condition is the noncommutative analogue of countable additivity for measures. Later, normal states and traces will provide the analytic data used to distinguish finite, semifinite, and type $III$ behaviour. ## The Course Roadmap What should a reader expect to know by the end of the course? The first part builds the concrete foundations: weak and strong [operator topologies](/page/Operator%20Topologies), von Neumann algebras, normal functionals, and monotone convergence. The second part proves the bicommutant theorem and uses it to organise examples through commutants, cyclic vectors, and reducing projections. The middle of the course studies projections and factors. Murray-von Neumann equivalence replaces rank, central projections decompose algebras into direct summands, and factors isolate the indecomposable cases. The type classification then separates finite type $I$ behaviour, type $II$ trace phenomena, and type $III$ algebras where no nonzero semifinite trace controls the projection theory. The final part connects the structure theory to representation and quantum language. Normal states describe expectation values, traces encode dimension-like information, and tensor products model composite systems. By the end, the reader should be able to move between concrete Hilbert-space operator calculations, commutant arguments, and the projection-theoretic vocabulary used in modern operator algebra. The introductory discussion has now assembled the basic language of operators, states, traces, and tensor products, so the next step is to make that language concrete on Hilbert space. Rather than focusing on norm approximation alone, the course now studies algebras through how their operators act on vectors and what closure properties that action reveals. # 1. Weak Operator Closure and Concrete Operator Algebras Building on the introductory weak-closure viewpoint, this chapter sets up the concrete viewpoint on von Neumann algebras: an algebra of bounded operators is studied through how its operators act on Hilbert space vectors, rather than through norm approximation alone. The prerequisites are the basic language of Hilbert spaces, bounded linear operators, adjoints, [weak convergence](/page/Weak%20Convergence), and the [uniform boundedness principle](/theorems/549). The guiding point is that weak and strong operator limits retain much more spectral and order structure than arbitrary norm limits. We begin by comparing the relevant topologies on $\mathcal{L}(H)$, then define von Neumann algebras as operator-topologically closed $*$-algebras, and finish by introducing normal functionals as the operational tests for ultraweak convergence. ## Operator Topologies on $\mathcal{L}(H)$ The first question is how to express convergence of bounded operators when norm convergence is too rigid. In many examples from spectral theory and multiplication operators, a sequence or net of operators converges on each vector, or only through each matrix coefficient, even though the operator norms do not converge. [definition: Strong Operator Topology] Let $H$ be a Hilbert space. For each $\xi\in H$, define a seminorm \begin{align*} p_\xi:\mathcal{L}(H)&\to [0,\infty), & p_\xi(T)&=\|T\xi\|_H. \end{align*} The strong operator topology on $\mathcal{L}(H)$ is the topology generated by the family $(p_\xi)_{\xi\in H}$. A net $(T_i)_{i \in I}$ converges strongly to $T \in \mathcal{L}(H)$ if $\|T_i\xi - T\xi\|_H \to 0$ for every $\xi \in H$. [/definition] Strong convergence is pointwise norm convergence on the Hilbert space, so it remembers the image of every fixed vector. It still misses phase-by-phase scalar information that may survive even when vectors do not converge in norm, and this leads to a still weaker topology built from matrix coefficients. [definition: Weak Operator Topology] Let $H$ be a Hilbert space. For each $\xi,\eta\in H$, define a seminorm \begin{align*} p_{\xi,\eta}:\mathcal{L}(H)&\to [0,\infty), & p_{\xi,\eta}(T)&=|(T\xi,\eta)_H|. \end{align*} The weak operator topology on $\mathcal{L}(H)$ is the topology generated by the family $(p_{\xi,\eta})_{\xi,\eta\in H}$. A net $(T_i)_{i \in I}$ converges weakly to $T \in \mathcal{L}(H)$ if $(T_i\xi,\eta)_H \to (T\xi,\eta)_H$ for every $\xi,\eta \in H$. [/definition] Weak operator convergence records all matrix coefficients, while strong convergence records actual vector limits. To use these topologies without ambiguity, we need the basic comparison theorem that places norm, strong, and weak convergence in a single hierarchy. [quotetheorem:9260] [citeproof:9260] The missing reverse implications are not minor technicalities; they explain why von Neumann algebras behave differently from $C^*$-algebras. The projections $P_N$ in the next example converge strongly to $I$ but stay distance $1$ from $I$ in operator norm, so strong convergence need not imply norm convergence. The powers of the unilateral shift in the following example converge weakly to $0$ but do not converge strongly, so weak operator convergence need not imply strong convergence. These counterexamples are not pathologies; they are the basic reason that different operator closures lead to genuinely different categories of operator algebra. The following example gives a standard strong limit with no norm convergence. [example: Finite Rank Projections Converge Strongly] Let $H=\ell^2(\mathbb N)$ with standard orthonormal basis $(e_n)_{n\ge 1}$, and let $P_N$ be the projection onto $\operatorname{span}\{e_1,\dots,e_N\}$. For $\xi=(\xi_n)_{n\ge 1}\in \ell^2(\mathbb N)$, the projection is \begin{align*} P_N\xi=(\xi_1,\dots,\xi_N,0,0,\dots). \end{align*} Hence \begin{align*} (I-P_N)\xi=(0,\dots,0,\xi_{N+1},\xi_{N+2},\dots). \end{align*} Taking the $\ell^2$ norm gives \begin{align*} \|(I-P_N)\xi\|_{\ell^2}^2=\sum_{n=N+1}^{\infty}|\xi_n|^2. \end{align*} Since $\xi\in\ell^2(\mathbb N)$, the series $\sum_{n=1}^{\infty}|\xi_n|^2$ converges, so its tails tend to $0$. Therefore $\|(I-P_N)\xi\|_{\ell^2}\to 0$, which means $P_N\xi\to \xi$ for every $\xi\in H$. Thus $P_N\to I$ strongly. The convergence is not norm convergence. For every $\xi\in\ell^2(\mathbb N)$, \begin{align*} \|(I-P_N)\xi\|_{\ell^2}^2=\sum_{n=N+1}^{\infty}|\xi_n|^2\le \sum_{n=1}^{\infty}|\xi_n|^2=\|\xi\|_{\ell^2}^2. \end{align*} So $\|I-P_N\|_{\mathcal L(H)}\le 1$. On the other hand, \begin{align*} (I-P_N)e_{N+1}=e_{N+1}. \end{align*} Since $\|e_{N+1}\|_{\ell^2}=1$, this gives $\|I-P_N\|_{\mathcal L(H)}\ge 1$. Hence \begin{align*} \|I-P_N\|_{\mathcal L(H)}=1 \end{align*} for every $N$. The finite-rank projections therefore converge strongly to $I$ while staying distance $1$ from $I$ in operator norm, so strong closure can contain operators that are not norm limits. [/example] Weak convergence is even less rigid. It is especially natural when the information available about an operator consists of transition amplitudes or matrix coefficients rather than vectors themselves. [example: Shifts Converge Weakly But Not Strongly] Let $S \in \mathcal L(\ell^2(\mathbb N))$ be the unilateral shift, so $Se_n=e_{n+1}$ and therefore $S^k e_n=e_{n+k}$. We show that $S^k\to 0$ in the weak operator topology but not in the strong operator topology. First take finitely supported vectors \begin{align*} \xi=\sum_{n=1}^m \xi_n e_n \quad \text{and} \quad \eta=\sum_{r=1}^q \eta_r e_r. \end{align*} Then \begin{align*} S^k\xi=\sum_{n=1}^m \xi_n e_{n+k}. \end{align*} If $k>q$, then $n+k>q$ for every $1\le n\le m$, so each basis vector $e_{n+k}$ is orthogonal to each basis vector appearing in $\eta$. Hence \begin{align*} (S^k\xi,\eta)_{\ell^2}=0. \end{align*} Thus $(S^k\xi,\eta)_{\ell^2}\to 0$ whenever both vectors are finitely supported. Now let $\xi,\eta\in\ell^2(\mathbb N)$ be arbitrary. Given $\varepsilon>0$, choose finitely supported $\xi^{(0)},\eta^{(0)}$ such that \begin{align*} \|\xi-\xi^{(0)}\|_{\ell^2}<\frac{\varepsilon}{3(\|\eta\|_{\ell^2}+1)} \end{align*} and \begin{align*} \|\eta-\eta^{(0)}\|_{\ell^2}<\frac{\varepsilon}{3(\|\xi^{(0)}\|_{\ell^2}+1)}. \end{align*} Since $S$ preserves the orthonormal basis, $\|S^k\zeta\|_{\ell^2}=\|\zeta\|_{\ell^2}$ for every $\zeta\in\ell^2(\mathbb N)$. For all sufficiently large $k$, the finitely supported case gives $(S^k\xi^{(0)},\eta^{(0)})_{\ell^2}=0$. For such $k$, \begin{align*} (S^k\xi,\eta)_{\ell^2}=(S^k(\xi-\xi^{(0)}),\eta)_{\ell^2}+(S^k\xi^{(0)},\eta-\eta^{(0)})_{\ell^2}. \end{align*} By Cauchy--Schwarz and $\|S^k\zeta\|_{\ell^2}=\|\zeta\|_{\ell^2}$, \begin{align*} |(S^k(\xi-\xi^{(0)}),\eta)_{\ell^2}|\le \|\xi-\xi^{(0)}\|_{\ell^2}\|\eta\|_{\ell^2}<\frac{\varepsilon}{3}. \end{align*} Also, \begin{align*} |(S^k\xi^{(0)},\eta-\eta^{(0)})_{\ell^2}|\le \|\xi^{(0)}\|_{\ell^2}\|\eta-\eta^{(0)}\|_{\ell^2}<\frac{\varepsilon}{3}. \end{align*} Therefore $|(S^k\xi,\eta)_{\ell^2}|<\varepsilon$ for all sufficiently large $k$, so $S^k\to 0$ weakly. The convergence is not strong. Indeed, \begin{align*} S^k e_1=e_{k+1}. \end{align*} Since $\|e_{k+1}\|_{\ell^2}=1$ for every $k$, \begin{align*} \|S^k e_1-0\|_{\ell^2}=1. \end{align*} Thus $S^k e_1$ does not converge to $0$ in norm, so $S^k$ does not converge strongly to $0$. This gives a concrete sequence that converges weakly but not strongly. [/example] The shift example shows that weak convergence alone may give little direct control over vector norms. Before using strong limits inside algebras, we need the uniform boundedness principle to guarantee that pointwise convergence of operators still imposes a global operator-norm bound on the approximating family. [quotetheorem:549] [citeproof:549] This boundedness result lets us move between weak information and strong information in special algebraic situations, but its hypotheses are doing real work. Pointwise convergence is essential: a family such as $nI$ is not uniformly bounded, and it also does not converge on any nonzero vector. Completeness is also essential through the uniform boundedness principle, which is why the theorem is stated on a Hilbert space rather than on an arbitrary [inner product space](/page/Inner%20Product%20Space). The theorem does not say that every pointwise bounded family converges, nor does it turn weak convergence into strong convergence by itself; it only supplies the norm bound needed before algebraic and self-adjoint structure can be used. The next section makes that extra structure precise for self-adjoint unital algebras. ## Von Neumann Algebras as Operator-Closed Algebras The next problem is to decide which concrete operator algebras should count as closed under the limiting processes visible on Hilbert space. Norm-closed $*$-algebras are $C^*$-algebras; von Neumann algebras are the concrete $*$-algebras closed under weak, or equivalently strong, operator limits. [definition: Concrete Von Neumann Algebra] Let $H$ be a Hilbert space. A concrete von Neumann algebra on $H$ is a subset $M\subset\mathcal{L}(H)$ together with the inclusion map \begin{align*} \iota_M:M&\to\mathcal{L}(H), & \iota_M(T)&=T, \end{align*} such that $M$ is closed in the weak operator topology and is closed under the operations \begin{align*} a_M:M\times M&\to M, & a_M(S,T)&=S+T. \end{align*} \begin{align*} s_M:\mathbb C\times M&\to M, & s_M(\lambda,T)&=\lambda T. \end{align*} \begin{align*} m_M:M\times M&\to M, & m_M(S,T)&=ST. \end{align*} \begin{align*} {*}_M:M&\to M, & {*}_M(T)&=T^*. \end{align*} \begin{align*} u_M:\mathbb C&\to M, & u_M(\lambda)&=\lambda I_H. \end{align*} [/definition] The definition is intentionally representation-dependent: $M$ is an algebra of actual operators on a specified Hilbert space. At first this raises a stability problem: the weak operator topology appears in the definition, while many operator-algebra constructions are more naturally checked in the strong operator topology. Since strong and weak closures differ for arbitrary sets of operators, one needs a special closure comparison for linear operator spaces before the two formulations of von Neumann algebra can be treated as equivalent. [quotetheorem:9261] [citeproof:9261] This theorem justifies saying that a von Neumann algebra is strongly closed as well as weakly closed. There are two related statements being used here. The statement needed for von Neumann algebras is the algebraic version above: a unital $*$-subalgebra has the same strong and weak operator closures. The proof, however, uses only the linear-subspace property: the set of simultaneous values $(a\xi_1,\dots,a\xi_n)$ is a linear subspace of $H^{\oplus n}$, so orthogonality against its complement can be tested by weak operator continuous linear functionals. Hence the sharper linear-topological result says that every linear subspace of $\mathcal{L}(H)$ has the same strong and weak operator closures. The unital and self-adjoint hypotheses are present in the theorem-as-used because von Neumann algebras are unital $*$-algebras, not because those hypotheses drive the closure comparison. The linearity requirement still matters. For arbitrary subsets, weak and strong closures can differ because the argument above needs a single approximant to work simultaneously on finitely many vectors and to remain in a linear set of possible simultaneous values. For instance, if $S$ is the unilateral shift on $\ell^2(\mathbb N)$, then $\{S^n:n\ge 1\}$ has $0$ in its weak operator closure, but $0$ is not in its strong operator closure because $\|S^n e_1\|_{\ell^2}=1$ for every $n$. The unit matters instead for the ambient algebraic convention: a nonunital corner such as $\mathbb C P\subset\mathcal{L}(H)$ for a proper projection $P\ne I$ may be weakly and strongly closed, but it is not a unital algebra on the Hilbert space $H$ with unit $I_H$. The definition deliberately records both the algebraic convention and the Hilbert space representation, while the closure theorem records a separate linear-topological fact. [example: The Algebra $\mathcal{L}(H)$] For every Hilbert space $H$, the full operator algebra $\mathcal{L}(H)$ is a concrete von Neumann algebra on $H$. The unit is $I_H$, and if $S,T\in\mathcal{L}(H)$ and $\lambda\in\mathbb C$, then $S+T$, $\lambda T$, $ST$, and $T^*$ are again bounded operators on $H$. Since the weak operator topology is a topology on the ambient set $\mathcal{L}(H)$ itself, the subset $\mathcal{L}(H)\subset\mathcal{L}(H)$ is closed: any weak operator limit of operators in the whole space is still an element of the whole space. If $H\ne\{0\}$, the commutant of $\mathcal{L}(H)$ is exactly $\mathbb C I_H$. Indeed, suppose $R\in\mathcal{L}(H)$ commutes with every operator in $\mathcal{L}(H)$. Choose a unit vector $y\in H$. For each $x\in H$, define the rank-one operator \begin{align*} \theta_{x,y}(z)=(z,y)_H x. \end{align*} Since $R\theta_{x,y}=\theta_{x,y}R$, evaluating at $y$ gives \begin{align*} R\theta_{x,y}y=\theta_{x,y}Ry. \end{align*} Now $\theta_{x,y}y=(y,y)_H x=x$, while $\theta_{x,y}Ry=(Ry,y)_H x$. Hence \begin{align*} Rx=(Ry,y)_H x. \end{align*} The scalar $\lambda=(Ry,y)_H$ is independent of $x$, so $R=\lambda I_H$. Conversely, every scalar multiple $\lambda I_H$ commutes with every bounded operator because \begin{align*} (\lambda I_H)T=\lambda T=T(\lambda I_H). \end{align*} Thus the largest concrete von Neumann algebra has the smallest possible commutant. [/example] The first family of substantive examples comes from diagonal operators. These are the operator-algebraic version of bounded scalar functions on a countable measure space. [example: Diagonal Algebra On $\ell^2(\mathbb N)$] For $a=(a_n)_{n\ge 1}\in \ell^\infty(\mathbb N)$, define $M_a$ first on basis vectors by $M_a e_n=a_n e_n$. If $\xi=(\xi_n)_{n\ge 1}\in \ell^2(\mathbb N)$, then \begin{align*} \|M_a\xi\|_{\ell^2}^2=\sum_{n=1}^{\infty}|a_n\xi_n|^2\le \|a\|_{\ell^\infty}^2\sum_{n=1}^{\infty}|\xi_n|^2=\|a\|_{\ell^\infty}^2\|\xi\|_{\ell^2}^2. \end{align*} Thus $M_a$ is bounded and $\|M_a\|_{\mathcal L(\ell^2)}\le \|a\|_{\ell^\infty}$. The set \begin{align*} D=\{M_a:a\in\ell^\infty(\mathbb N)\} \end{align*} is a unital $*$-subalgebra. Indeed, for $a,b\in\ell^\infty(\mathbb N)$, $\lambda\in\mathbb C$, and every basis vector $e_n$, \begin{align*} (M_a+M_b)e_n=M_a e_n+M_b e_n=a_n e_n+b_n e_n=(a_n+b_n)e_n=M_{a+b}e_n. \end{align*} Also \begin{align*} (\lambda M_a)e_n=\lambda a_n e_n=M_{\lambda a}e_n. \end{align*} For multiplication, \begin{align*} M_aM_b e_n=M_a(b_n e_n)=b_n a_n e_n=M_{ab}e_n. \end{align*} The constant sequence $1=(1,1,\dots)$ gives $M_1=I$. Finally, since \begin{align*} (M_a e_m,e_n)_{\ell^2}=a_m\delta_{mn} \end{align*} and \begin{align*} (e_m,M_{\bar a}e_n)_{\ell^2}=a_n\delta_{mn}=a_m\delta_{mn}, \end{align*} we have $M_a^*=M_{\bar a}$. Hence $D$ is closed under addition, scalar multiplication, products, adjoints, and contains the identity. It remains to check weak operator closedness. Suppose $(M_{a^{(i)}})_{i\in I}$ is a net in $D$ converging weakly to $T\in\mathcal L(\ell^2(\mathbb N))$. For each fixed $m,n\ge 1$, \begin{align*} (Te_m,e_n)_{\ell^2}=\lim_i(M_{a^{(i)}}e_m,e_n)_{\ell^2}=\lim_i a^{(i)}_m\delta_{mn}. \end{align*} If $m\ne n$, this gives $(Te_m,e_n)_{\ell^2}=0$. Define \begin{align*} a_m=(Te_m,e_m)_{\ell^2}. \end{align*} Then $|a_m|\le \|T\|_{\mathcal L(\ell^2)}\|e_m\|_{\ell^2}^2=\|T\|_{\mathcal L(\ell^2)}$, so $a=(a_m)_{m\ge 1}\in\ell^\infty(\mathbb N)$. The coefficient computation gives \begin{align*} (Te_m,e_n)_{\ell^2}=(a_m e_m,e_n)_{\ell^2} \end{align*} for every $n$. Since the standard basis determines vectors in $\ell^2(\mathbb N)$, $Te_m=a_m e_m=M_a e_m$ for every $m$. By linearity and continuity, $T=M_a$ on all of $\ell^2(\mathbb N)$. Therefore $D$ is weakly closed, hence it is a concrete von Neumann algebra naturally identified with $\ell^\infty(\mathbb N)$. [/example] The diagonal example should be read as multiplication by functions on a discrete measure space. Passing to a general measure space changes the notation but not the operator-theoretic mechanism: basis vectors are replaced by $L^2$ functions, diagonal entries are replaced by essentially bounded functions, and matrix coefficients become integrals. This viewpoint also explains why commutativity persists in the next example, since multiplication operators commute pointwise before passing to equivalence classes. [example: Multiplication Algebra $L^\infty(X,\mu)$] Let $(X,\mathcal E,\mu)$ be a $\sigma$-finite measure space and let $H=L^2(X,\mu)$. For $f\in L^\infty(X,\mu)$, define $M_f g=fg$. This is bounded because, for every $g\in L^2(X,\mu)$, \begin{align*} \|M_f g\|_{L^2}^2=\int_X |f g|^2\,d\mu\le \|f\|_{L^\infty}^2\int_X |g|^2\,d\mu=\|f\|_{L^\infty}^2\|g\|_{L^2}^2. \end{align*} Thus $\|M_f\|\le \|f\|_{L^\infty}$. The multiplication operators form a commutative unital $*$-subalgebra: for $f,k\in L^\infty(X,\mu)$, $\lambda\in\mathbb C$, and $g\in L^2(X,\mu)$, \begin{align*} (M_f+M_k)g=fg+kg=(f+k)g=M_{f+k}g. \end{align*} \begin{align*} (\lambda M_f)g=\lambda fg=(\lambda f)g=M_{\lambda f}g. \end{align*} \begin{align*} M_fM_k g=f(kg)=(fk)g=M_{fk}g=M_{kf}g=M_kM_f g. \end{align*} The constant function $1$ gives $M_1=I$. Also, \begin{align*} (M_f g,h)_{L^2}=\int_X f g\bar h\,d\mu=\int_X g\,\overline{\bar f h}\,d\mu=(g,M_{\bar f}h)_{L^2}, \end{align*} so $M_f^*=M_{\bar f}$. It remains to prove weak operator closedness. Suppose $M_{f_i}\to T$ weakly in $\mathcal L(L^2(X,\mu))$. For every measurable set $E$, write $P_E=M_{\mathbf 1_E}$. Since each $M_{f_i}$ commutes with $P_E$, for $g,h\in L^2(X,\mu)$ we have \begin{align*} (P_ETg,h)_{L^2}=(Tg,P_Eh)_{L^2}=\lim_i(M_{f_i}g,P_Eh)_{L^2}. \end{align*} Because $P_E=P_E^*$ and $P_EM_{f_i}=M_{f_i}P_E$, \begin{align*} (M_{f_i}g,P_Eh)_{L^2}=(P_EM_{f_i}g,h)_{L^2}=(M_{f_i}P_Eg,h)_{L^2}. \end{align*} Taking the weak limit again gives \begin{align*} (P_ETg,h)_{L^2}=(TP_Eg,h)_{L^2}. \end{align*} Since this holds for all $h$, $P_ETg=TP_Eg$ for every $g$, so $T$ commutes with every multiplication projection $P_E$. Choose measurable sets $X_m$ with $\mu(X_m)<\infty$, $X_m\subset X_{m+1}$, and $\bigcup_m X_m=X$ up to a null set. Put $u_m=\mathbf 1_{X_m}$ and define $a_m=Tu_m$ on $X_m$. If $A\subset X_m$ is measurable, then \begin{align*} T\mathbf 1_A=TP_Au_m=P_ATu_m=\mathbf 1_A a_m. \end{align*} By linearity, if $s=\sum_{r=1}^q c_r\mathbf 1_{A_r}$ is a [simple function](/page/Simple%20Function) supported in $X_m$, then \begin{align*} Ts=\sum_{r=1}^q c_rT\mathbf 1_{A_r}=\sum_{r=1}^q c_r\mathbf 1_{A_r}a_m=a_m s. \end{align*} The operator bound gives \begin{align*} \|a_m s\|_{L^2}=\|Ts\|_{L^2}\le \|T\|\|s\|_{L^2}. \end{align*} Taking $s=\mathbf 1_A$ shows $\int_A |a_m|^2\,d\mu\le \|T\|^2\mu(A)$ for every $A\subset X_m$ of finite measure, so $|a_m|\le \|T\|$ almost everywhere on $X_m$. If $m\le n$, then on $X_m$, \begin{align*} a_n\mathbf 1_{X_m}=P_{X_m}Tu_n=TP_{X_m}u_n=Tu_m=a_m. \end{align*} Hence the functions $a_m$ agree on overlaps. Define $f$ by $f=a_m$ on $X_m$; then $f\in L^\infty(X,\mu)$ and $\|f\|_{L^\infty}\le \|T\|$. For every simple function $s$ supported in some $X_m$, the computation above gives $Ts=fs=M_fs$. Such simple functions are dense in $L^2(X,\mu)$, and both $T$ and $M_f$ are bounded, so $T=M_f$ on all of $L^2(X,\mu)$. Therefore the multiplication algebra is weakly closed, hence it is a commutative von Neumann algebra on $L^2(X,\mu)$. [/example] Multiplication algebras give the commutative model: every operator acts by scaling fibres rather than mixing them. A first noncommutative variation keeps the Hilbert space decomposed but allows arbitrary operators inside each summand. The only forbidden behaviour is movement between the summands, and weak operator limits preserve that prohibition because off-diagonal matrix coefficients remain zero in the limit. [example: Block Diagonal Matrix Algebras] Let $H=H_1\oplus H_2$, and write vectors as pairs $(\xi_1,\xi_2)$ with $\xi_r\in H_r$. Define \begin{align*} M=\{T_{A,B}:A\in\mathcal L(H_1),\ B\in\mathcal L(H_2)\}, \end{align*} where \begin{align*} T_{A,B}(\xi_1,\xi_2)=(A\xi_1,B\xi_2). \end{align*} For such an operator, \begin{align*} \|T_{A,B}(\xi_1,\xi_2)\|_H^2=\|A\xi_1\|_{H_1}^2+\|B\xi_2\|_{H_2}^2\le \max\{\|A\|^2,\|B\|^2\}(\|\xi_1\|_{H_1}^2+\|\xi_2\|_{H_2}^2), \end{align*} so $T_{A,B}\in\mathcal L(H)$. The set $M$ is a unital $*$-subalgebra. For $A,C\in\mathcal L(H_1)$, $B,D\in\mathcal L(H_2)$, and $\lambda\in\mathbb C$, \begin{align*} (T_{A,B}+T_{C,D})(\xi_1,\xi_2)=((A+C)\xi_1,(B+D)\xi_2)=T_{A+C,B+D}(\xi_1,\xi_2). \end{align*} Also, \begin{align*} (\lambda T_{A,B})(\xi_1,\xi_2)=(\lambda A\xi_1,\lambda B\xi_2)=T_{\lambda A,\lambda B}(\xi_1,\xi_2). \end{align*} Multiplication is blockwise: \begin{align*} T_{A,B}T_{C,D}(\xi_1,\xi_2)=T_{A,B}(C\xi_1,D\xi_2)=(AC\xi_1,BD\xi_2)=T_{AC,BD}(\xi_1,\xi_2). \end{align*} The identity is $T_{I_{H_1},I_{H_2}}$. Finally, for $\xi=(\xi_1,\xi_2)$ and $\eta=(\eta_1,\eta_2)$, \begin{align*} (T_{A,B}\xi,\eta)_H=(A\xi_1,\eta_1)_{H_1}+(B\xi_2,\eta_2)_{H_2}=(\xi_1,A^*\eta_1)_{H_1}+(\xi_2,B^*\eta_2)_{H_2}=(\xi,T_{A^*,B^*}\eta)_H, \end{align*} so $T_{A,B}^*=T_{A^*,B^*}$. It remains to check weak operator closedness. Suppose $T_{A_i,B_i}\to T$ weakly in $\mathcal L(H)$. If $\xi_1\in H_1$ and $\eta_2\in H_2$, then \begin{align*} (T(\xi_1,0),(0,\eta_2))_H=\lim_i(T_{A_i,B_i}(\xi_1,0),(0,\eta_2))_H=\lim_i(A_i\xi_1,0)_{H}=0. \end{align*} Since this holds for every $\eta_2\in H_2$, the $H_2$-component of $T(\xi_1,0)$ is zero. Thus $T(H_1\oplus 0)\subset H_1\oplus 0$. Similarly, if $\xi_2\in H_2$ and $\eta_1\in H_1$, then \begin{align*} (T(0,\xi_2),(\eta_1,0))_H=\lim_i(T_{A_i,B_i}(0,\xi_2),(\eta_1,0))_H=\lim_i(0,B_i\xi_2)_{H}=0, \end{align*} so $T(0\oplus H_2)\subset 0\oplus H_2$. Define $A\in\mathcal L(H_1)$ and $B\in\mathcal L(H_2)$ by \begin{align*} T(\xi_1,0)=(A\xi_1,0) \end{align*} and \begin{align*} T(0,\xi_2)=(0,B\xi_2). \end{align*} These maps are bounded because they are restrictions of the bounded operator $T$. For every $(\xi_1,\xi_2)\in H$, \begin{align*} T(\xi_1,\xi_2)=T(\xi_1,0)+T(0,\xi_2)=(A\xi_1,B\xi_2)=T_{A,B}(\xi_1,\xi_2). \end{align*} Hence $T\in M$, so $M$ is weakly closed. Therefore the block diagonal algebra is a concrete von Neumann algebra on $H$, and the central projections onto $H_1$ and $H_2$ record the decomposition into the two summands. [/example] These examples illustrate the basic dictionary: von Neumann algebras are algebras of observables stable under the limiting operations that preserve all vector-state measurements. The next order-theoretic result is one of the main technical advantages of this closure. [quotetheorem:9262] [citeproof:9262] Monotone convergence is a major point of contact between measure theory and von Neumann algebras. The boundedness hypothesis cannot simply be dropped: the increasing sequence $nI$ has no bounded operator as a strong limit. Monotonicity is also essential, since a bounded sequence of [self-adjoint operators](/page/Self-Adjoint%20Operators) can oscillate and fail to converge strongly, for instance $(-1)^nI$. Self-adjointness is what gives an order relation and positive square-root estimates; without it, there is no comparable monotone order on arbitrary operators. Thus the theorem is an order-completeness statement for bounded self-adjoint parts of von Neumann algebras, not a [compactness theorem](/theorems/2748) for arbitrary bounded families. It says that bounded increasing approximation by observables has an observable limit inside the same algebra. ## Normal Functionals and Ultraweak Convergence The final question in this chapter is how to formulate convergence using the linear functionals that arise from states, traces, and density operators. Weak operator convergence tests only vector matrix coefficients; ultraweak convergence tests the larger class of normal linear functionals. [definition: Vector Functional] Let $H$ be a Hilbert space and let $\xi,\eta\in H$. The vector functional $\omega_{\xi,\eta}:\mathcal{L}(H)\to\mathbb C$ is defined by \begin{align*} \omega_{\xi,\eta}(T)=(T\xi,\eta)_H. \end{align*} [/definition] Vector functionals are the basic weak operator tests, and they already encode the weak operator topology. For states, traces, and statistical mixtures, however, we also need countable absolutely summable combinations of such matrix coefficients; these are the functionals compatible with monotone limits and countable measurement procedures. [definition: Normal Functional On $\mathcal{L}(H)$] A linear functional $\varphi:\mathcal{L}(H)\to\mathbb C$ is normal if there exist sequences $(\xi_n)_{n\ge 1}$ and $(\eta_n)_{n\ge 1}$ in $H$ such that \begin{align*} \sum_{n=1}^{\infty}\|\xi_n\|_H\|\eta_n\|_H<\infty \end{align*} and, for each $n\ge 1$, the vector functional \begin{align*} \omega_{\xi_n,\eta_n}:\mathcal{L}(H)&\to\mathbb C, & \omega_{\xi_n,\eta_n}(T)&=(T\xi_n,\eta_n)_H \end{align*} satisfies \begin{align*} \varphi(T)=\sum_{n=1}^{\infty}\omega_{\xi_n,\eta_n}(T) \end{align*} for every $T\in\mathcal{L}(H)$. [/definition] This definition is equivalent to saying that normal functionals on $\mathcal{L}(H)$ are represented by trace-class operators. If $M\subset\mathcal{L}(H)$ is a von Neumann algebra, a functional on $M$ is called normal when it is the restriction to $M$ of a normal functional on $\mathcal{L}(H)$; equivalently, it is continuous for the ultraweak topology induced from $\mathcal{L}(H)$. The vector-functional form is better suited to the present discussion because it connects directly to operator-topological convergence, and it naturally defines a topology by asking all normal functionals to converge. [definition: Ultraweak Topology] Let $H$ be a Hilbert space. The ultraweak topology on $\mathcal{L}(H)$ is the topology generated by all normal functionals on $\mathcal{L}(H)$. A net $(T_i)_{i\in I}$ converges ultraweakly to $T$ if $\varphi(T_i)\to\varphi(T)$ for every normal functional $\varphi$. [/definition] Ultraweak convergence is stronger than weak operator convergence because every vector functional is normal. On bounded sets, the two topologies are closely related for many practical purposes, but normal functionals are the correct dual objects for integration, states, and monotone limits. [example: Ultraweak Convergence Of Diagonal Operators] Let $D=\{M_a:a\in\ell^\infty(\mathbb N)\}$ act diagonally on $\ell^2(\mathbb N)$, so $M_a e_n=a_n e_n$. If $\varphi$ is the restriction to $D$ of a normal functional on $\mathcal L(\ell^2(\mathbb N))$, then there are vectors $(\xi^{(j)})_{j\ge 1}$ and $(\eta^{(j)})_{j\ge 1}$ in $\ell^2(\mathbb N)$ such that \begin{align*} \sum_{j=1}^{\infty}\|\xi^{(j)}\|_{\ell^2}\|\eta^{(j)}\|_{\ell^2}<\infty \end{align*} and \begin{align*} \varphi(M_a)=\sum_{j=1}^{\infty}(M_a\xi^{(j)},\eta^{(j)})_{\ell^2}. \end{align*} Writing $\xi^{(j)}=(\xi^{(j)}_n)_{n\ge 1}$ and $\eta^{(j)}=(\eta^{(j)}_n)_{n\ge 1}$, each summand is \begin{align*} (M_a\xi^{(j)},\eta^{(j)})_{\ell^2}=\sum_{n=1}^{\infty}a_n\xi^{(j)}_n\overline{\eta^{(j)}_n}. \end{align*} The absolute convergence estimate \begin{align*} \sum_{j=1}^{\infty}\sum_{n=1}^{\infty}|a_n\xi^{(j)}_n\overline{\eta^{(j)}_n}|\le \|a\|_{\ell^\infty}\sum_{j=1}^{\infty}\|\xi^{(j)}\|_{\ell^2}\|\eta^{(j)}\|_{\ell^2} \end{align*} uses Cauchy--Schwarz for each fixed $j$. Hence the order of summation may be exchanged, and with \begin{align*} b_n=\sum_{j=1}^{\infty}\xi^{(j)}_n\overline{\eta^{(j)}_n} \end{align*} we get \begin{align*} \varphi(M_a)=\sum_{n=1}^{\infty}a_n b_n. \end{align*} Moreover, \begin{align*} \sum_{n=1}^{\infty}|b_n|\le \sum_{j=1}^{\infty}\sum_{n=1}^{\infty}|\xi^{(j)}_n\overline{\eta^{(j)}_n}|\le \sum_{j=1}^{\infty}\|\xi^{(j)}\|_{\ell^2}\|\eta^{(j)}\|_{\ell^2}<\infty, \end{align*} so $b\in\ell^1(\mathbb N)$. Conversely, if $b\in\ell^1(\mathbb N)$, define vectors $\xi^{(n)},\eta^{(n)}\in\ell^2(\mathbb N)$ by $\xi^{(n)}=\sqrt{|b_n|}e_n$ and, when $b_n\ne 0$, \begin{align*} \eta^{(n)}=\frac{\overline{b_n}}{\sqrt{|b_n|}}e_n, \end{align*} while $\eta^{(n)}=0$ when $b_n=0$. Then \begin{align*} \sum_{n=1}^{\infty}\|\xi^{(n)}\|_{\ell^2}\|\eta^{(n)}\|_{\ell^2}=\sum_{n=1}^{\infty}|b_n|<\infty \end{align*} and \begin{align*} (M_a\xi^{(n)},\eta^{(n)})_{\ell^2}=a_n b_n. \end{align*} Thus $a\mapsto\sum_{n=1}^{\infty}a_nb_n$ is normal on $D$. Therefore a bounded net $M_{a^{(i)}}$ converges ultraweakly to $M_a$ precisely when, for every $b\in\ell^1(\mathbb N)$, \begin{align*} \sum_{n=1}^{\infty}a^{(i)}_n b_n\to \sum_{n=1}^{\infty}a_n b_n. \end{align*} In this diagonal case, ultraweak convergence is exactly weak-* convergence of $\ell^\infty(\mathbb N)$ tested against its predual $\ell^1(\mathbb N)$. [/example] The diagonal example shows how ultraweak convergence turns bounded pointwise limits into convergence against countably additive tests. We now need the theorem that connects this testing procedure to order: positive normal functionals should treat bounded increasing operator limits the way integrals treat increasing limits of functions. [quotetheorem:9263] [citeproof:9263] This theorem is the noncommutative analogue of monotone convergence for integrals, and its hypotheses mirror the measure-theoretic ones. Positivity is needed because the notation $\varphi(T_i)\uparrow\varphi(T)$ is an order statement; for a non-positive complex functional there is no monotone scalar sequence to compare. Normality is the continuity assumption that rules out singular functionals, just as countable additivity rules out merely finitely additive charges in measure theory. On $D\cong \ell^\infty(\mathbb N)$, a singular state extending the limit functional on convergent sequences satisfies $\psi(P_N)=0$ for the finite-rank diagonal projections $P_N$ while $\psi(I)=1$, even though $P_N\uparrow I$ strongly. Thus bounded monotone convergence detects the difference between normal $\ell^1$ tests and singular finitely additive tests. Boundedness and monotonicity are inherited from the operator [monotone convergence theorem](/theorems/509), so the result does not assert ultraweak continuity for arbitrary unbounded or non-monotone nets. It is one of the reasons normality is treated as part of the structure rather than as an optional regularity condition. [remark: What This Chapter Establishes] The chapter has introduced three levels of convergence on $\mathcal{L}(H)$: norm, strong operator, and weak operator convergence, together with ultraweak convergence from normal functionals. Von Neumann algebras are exactly the concrete unital $*$-algebras stable under the strong and weak operator limits relevant to Hilbert space measurements. The next chapter turns this closure condition into an algebraic criterion through commutants and the bicommutant theorem. [/remark] Weak and strong operator closure give the analytic shape of a von Neumann algebra, but they do not yet explain why commutants capture so much structure. The next chapter turns that closure picture into an algebraic one by showing that an algebra is determined by the operators that commute with it, and by the double commutant they generate. # 2. Commutants and the Bicommutant Theorem This chapter moves from the operator topologies of the first chapter to the algebraic operation that makes von Neumann algebras rigid: taking commutants. Continuing the weak and strong closure results from Chapter 1, the central question is how much of a concrete operator algebra is remembered by the operators that commute with it. The answer, the bicommutant theorem, says that for unital self-adjoint algebras on a Hilbert space, weak operator closure is the same operation as taking the double commutant. ## Operators Detected by What They Commute With The first problem is to turn the informal phrase "operators commuting with an algebra" into an object stable under the operations of operator theory. If $M \subseteq \mathcal{L}(H)$ is a set of bounded operators on a Hilbert space $H$, then the operators commuting with every element of $M$ form an algebra with strong closure properties. [definition: Commutant] Let $H$ be a Hilbert space and let $M \subseteq \mathcal{L}(H)$. The commutant of $M$ is \begin{align*} M' := \{T \in \mathcal{L}(H) : TS = ST \text{ for all } S \in M\}. \end{align*} The double commutant of $M$ is $M'' := (M')'$. [/definition] The commutant reverses inclusions: a larger family of test operators imposes more commutation relations. It also forgets redundant generators, since commuting with a set is the same as commuting with the unital algebra generated by that set. [example: Single Diagonal Operator With Simple Spectrum] Let $H=\ell^2(\mathbb N)$ with standard orthonormal basis $(e_n)$, and let $D e_n=\lambda_n e_n$, where $(\lambda_n)$ is bounded and the numbers $\lambda_n$ are pairwise distinct. Since $(\lambda_n)$ is bounded, $D$ is a bounded diagonal operator. Let $T\in\mathcal L(H)$ and write its matrix coefficients as $t_{ij}=(T e_j,e_i)_H$. If $TD=DT$, then applying both sides to $e_j$ gives \begin{align*} TDe_j=T(\lambda_j e_j)=\lambda_j T e_j. \end{align*} On the other hand, \begin{align*} DTe_j=D\left(\sum_i t_{ij}e_i\right)=\sum_i \lambda_i t_{ij}e_i. \end{align*} Taking the [inner product](/page/Inner%20Product) with $e_i$ in the equality $TDe_j=DTe_j$ gives \begin{align*} \lambda_j t_{ij}=\lambda_i t_{ij}. \end{align*} Thus \begin{align*} (\lambda_i-\lambda_j)t_{ij}=0. \end{align*} If $i\neq j$, then $\lambda_i\neq\lambda_j$, so $t_{ij}=0$. Therefore every operator commuting with $D$ is diagonal in the basis $(e_n)$. Conversely, if $T$ is any bounded diagonal operator, say $T e_n=a_n e_n$ with $(a_n)$ bounded, then for each basis vector $e_n$, \begin{align*} TDe_n=T(\lambda_n e_n)=\lambda_n a_n e_n. \end{align*} Also, \begin{align*} DTe_n=D(a_n e_n)=a_n\lambda_n e_n. \end{align*} Since $\lambda_n a_n=a_n\lambda_n$, we have $TDe_n=DTe_n$ for every $n$, hence $TD=DT$ on the dense linear span of the basis and therefore on all of $H$ by boundedness. Thus $\{D\}'$ is exactly the algebra of all [bounded diagonal operators](/theorems/4935). [/example] This example shows that commutation can recover spectral decompositions. The obstruction is that a commutant is defined by infinitely many equations $TS=ST$, so it is not immediately clear whether it is closed under limits or under the algebraic operations needed for a von Neumann algebra. The following structural result verifies that these equations are rigid enough to produce a closed operator algebra, with self-adjointness available when the original family is self-adjoint. [quotetheorem:9264] [citeproof:9264] The theorem explains why commutants already live on the von Neumann side of the theory: no closure operation has to be added after taking a commutant. The self-adjointness hypothesis is needed only for the final assertion; for instance, if $M$ consists of the unilateral shift on $\ell^2(\mathbb N)$, then $M'$ contains that shift but is not closed under adjoints. The result does not say that the original set $M$ is weakly closed or self-adjoint, only that the family of operators commuting with $M$ has those closure properties when the relevant hypothesis is present. To understand the geometry encoded by a commutant, we next compare commutation with invariance of subspaces. [definition: Invariant And Reducing Subspace] Let $M \subseteq \mathcal{L}(H)$ and let $K \subseteq H$ be a closed subspace. The subspace $K$ is invariant for $M$ if $T K \subseteq K$ for every $T \in M$. The subspace $K$ reduces $M$ if both $K$ and $K^\perp$ are invariant for $M$. [/definition] A closed subspace is represented by its [orthogonal projection](/theorems/437), so reducing subspaces should be visible inside an operator algebra rather than only in the geometry of $H$. The question is whether the projection onto a reducing subspace is exactly the same information as the reducing subspace itself. For self-adjoint families, the answer is yes: reducing means that every operator is block diagonal, and block diagonal operators commute with the corresponding coordinate projection. [quotetheorem:8394] [citeproof:8394] This projection criterion is one of the main ways in which von Neumann algebras are more geometric than norm-closed algebras. The self-adjointness assumption is essential: for the unilateral shift $S$ on $\ell^2(\mathbb N)$, the subspace spanned by $e_2,e_3,\dots$ is invariant for $S$, but its orthogonal complement is not invariant, and the corresponding projection does not commute with $S$. Thus invariant subspaces alone are not encoded by projections in the commutant; the theorem identifies reducing subspaces, where both halves of the [orthogonal decomposition](/theorems/436) are respected. This distinction is what makes projections in the commutant identify decompositions of the representation into independent parts. [example: Block Matrix Commutant] Identify $H=\mathbb C^n\otimes K$ with $K^n$ by writing vectors as sums of $e_j\otimes \eta$. Every $T\in \mathcal L(H)$ has block operators $T_{ij}\in\mathcal L(K)$ determined by \begin{align*} T(e_j\otimes \eta)=\sum_{i=1}^n e_i\otimes T_{ij}\eta. \end{align*} We compute the commutation relations with the matrix units $E_{ab}\otimes I_K$. For $1\leq a,b,j\leq n$ and $\eta\in K$, \begin{align*} T(E_{ab}\otimes I_K)(e_j\otimes \eta)=\delta_{bj}T(e_a\otimes \eta)=\delta_{bj}\sum_{i=1}^n e_i\otimes T_{ia}\eta. \end{align*} On the other hand, \begin{align*} (E_{ab}\otimes I_K)T(e_j\otimes \eta)=\sum_{i=1}^n (E_{ab}e_i)\otimes T_{ij}\eta=e_a\otimes T_{bj}\eta. \end{align*} If $T$ commutes with every $E_{ab}\otimes I_K$, these two expressions are equal for all $a,b,j,\eta$. Taking $j\neq b$ gives \begin{align*} 0=e_a\otimes T_{bj}\eta, \end{align*} so $T_{bj}=0$ whenever $b\neq j$. Since $b$ and $j$ were arbitrary, all off-diagonal blocks of $T$ are zero. Taking $j=b$ gives \begin{align*} \sum_{i=1}^n e_i\otimes T_{ia}\eta=e_a\otimes T_{bb}\eta. \end{align*} Comparing the coefficient of $e_a$ gives $T_{aa}\eta=T_{bb}\eta$ for every $\eta\in K$, hence $T_{aa}=T_{bb}$. Thus all diagonal blocks are one operator $A\in\mathcal L(K)$, and $T=I_n\otimes A$. Conversely, if $T=I_n\otimes A$, then for every $a,b,j,\eta$, \begin{align*} (I_n\otimes A)(E_{ab}\otimes I_K)(e_j\otimes \eta)=\delta_{bj}e_a\otimes A\eta. \end{align*} Also, \begin{align*} (E_{ab}\otimes I_K)(I_n\otimes A)(e_j\otimes \eta)=\delta_{bj}e_a\otimes A\eta. \end{align*} So $I_n\otimes A$ commutes with every $E_{ab}\otimes I_K$, and therefore with every element of $M_n(\mathbb C)\otimes I_K$ by linearity. Hence \begin{align*} (M_n(\mathbb C)\otimes I_K)'=I_n\otimes \mathcal L(K). \end{align*} The commutant records exactly the multiplicity space $K$: it removes the visible $M_n(\mathbb C)$ matrix action and leaves arbitrary operators acting uniformly on each copy of $K$. [/example] ## Cyclic And Separating Vectors The next problem is to detect an operator algebra through its action on vectors. Since weak operator convergence is defined by matrix coefficients, vectors that generate enough of the Hilbert space allow algebraic identities to be tested by scalar-valued functions. [definition: Cyclic Vector] Let $M \subseteq \mathcal{L}(H)$ be a unital algebra. A vector $\xi \in H$ is cyclic for $M$ if \begin{align*} \overline{M\xi}=H, \end{align*} where $M\xi:=\{T\xi:T\in M\}$. [/definition] Cyclicity says that the representation has a single vector from which the whole Hilbert space can be reached by applying the algebra. This is the operator-algebraic analogue of a module generated by one element. [example: Cyclic Vector For Diagonal Multiplication] Let $M=\ell^\infty(\mathbb N)$ act on $H=\ell^2(\mathbb N)$ by diagonal multiplication, so for $a=(a_n)\in \ell^\infty(\mathbb N)$ and $\xi=(\xi_n)\in \ell^2(\mathbb N)$, \begin{align*} (a\xi)_n=a_n\xi_n. \end{align*} We show that $\xi$ is cyclic for $M$ exactly when $\xi_n\neq 0$ for every $n$. First suppose $\xi_n\neq 0$ for every $n$. Fix $k\in\mathbb N$ and define $a^{(k)}\in\ell^\infty(\mathbb N)$ by $a^{(k)}_k=1/\xi_k$ and $a^{(k)}_n=0$ for $n\neq k$. This sequence is bounded because it has only one nonzero coordinate. Then \begin{align*} (a^{(k)}\xi)_k=a^{(k)}_k\xi_k=(1/\xi_k)\xi_k=1. \end{align*} For $n\neq k$, \begin{align*} (a^{(k)}\xi)_n=a^{(k)}_n\xi_n=0\cdot \xi_n=0. \end{align*} Hence $a^{(k)}\xi=e_k$, so every basis vector $e_k$ lies in $M\xi$. Therefore every finitely supported vector lies in $M\xi$ by linearity, and the finitely supported vectors are dense in $\ell^2(\mathbb N)$. Thus $\overline{M\xi}=H$, so $\xi$ is cyclic. Conversely, suppose $\xi_k=0$ for some $k$. For every $a\in\ell^\infty(\mathbb N)$, \begin{align*} (a\xi)_k=a_k\xi_k=a_k\cdot 0=0. \end{align*} Thus every vector in $M\xi$ has $k$th coordinate equal to $0$, and the same remains true for every vector in $\overline{M\xi}$ because coordinate evaluation is continuous on $\ell^2(\mathbb N)$. Since $(e_k)_k=1$, we have $e_k\notin\overline{M\xi}$, so $\overline{M\xi}\neq H$. Therefore $\xi$ is not cyclic. This identifies cyclicity for diagonal multiplication with having no zero coordinates. [/example] The dual notion asks whether a vector can witness that an operator is zero. This is naturally attached to an algebra because the vanishing is tested after applying all operators in that algebra. [definition: Separating Vector] Let $M \subseteq \mathcal{L}(H)$ be a unital algebra. A vector $\xi \in H$ is separating for $M$ if $T\xi=0$ for $T\in M$ implies $T=0$. [/definition] Cyclicity supplies many test vectors, while separation turns a single zero vector equation into a zero operator equation. The reason these two notions are paired in von Neumann algebra theory is that the commutant moves operators across the cyclic orbit without changing the generated subspace. The next theorem makes this precise and is the standard way to convert density of $M\xi$ into uniqueness for operators in $M'$. [quotetheorem:9265] [citeproof:9265] This theorem turns density questions into kernel questions. The unital hypothesis ensures that $\xi\in \overline{M\xi}$ and that the cyclic orbit is the right closed subspace to test; without it, degenerate algebras such as $M=\{0\}$ make cyclicity and separation unrelated. Self-adjointness is also doing real work: it turns the closed invariant subspace $\overline{M\xi}$ into a reducing subspace, so that its projection lies in $M'$. The theorem does not say that a cyclic vector for $M$ is separating for $M$ itself, and finite-dimensional diagonal algebras show these are different conditions. It is especially useful in matrix coefficient arguments, because separating vectors let us promote equality on a vector to equality of operators. [explanation: Matrix Coefficient Strategy] A typical argument begins with a proposed limit operator $T$ and a family of approximating operators $A_i\in M$. To prove weak operator convergence $A_i\to T$, it is enough to prove \begin{align*} (A_i\xi,\eta)_H \to (T\xi,\eta)_H \end{align*} for enough pairs of vectors $\xi,\eta$ whose linear span is dense in $H$. Cyclic vectors provide such dense families, often of the form $S\xi$ with $S\in M$ or $S\in M'$. Commutation then moves operators from one side of the inner product to the other, converting algebraic commutation relations into convergence of scalar coefficients. [/explanation] The argument just described is the engine behind the bicommutant theorem. Before stating it, we record a second canonical class of examples where the commutant has a concrete interpretation. [example: Left And Right Regular Representations] [claim]For a discrete group $G$ on $H=\ell^2(G)$, the operators $\lambda(g)$ commute with the operators $\rho(k)$, and the commutant of the von Neumann algebra generated by $\lambda(G)$ is the von Neumann algebra generated by $\rho(G)$.[/claim] [proof]For $\xi\in \ell^2(G)$ and $h,g,k\in G$, \begin{align*} (\lambda(g)\rho(k)\xi)(h)=(\rho(k)\xi)(g^{-1}h)=\xi(g^{-1}hk). \end{align*} Also, \begin{align*} (\rho(k)\lambda(g)\xi)(h)=(\lambda(g)\xi)(hk)=\xi(g^{-1}hk). \end{align*} Thus $\lambda(g)\rho(k)=\rho(k)\lambda(g)$ for all $g,k\in G$, so the von Neumann algebra generated by $\rho(G)$ is contained in the commutant of the von Neumann algebra generated by $\lambda(G)$. For the reverse inclusion, let $T$ commute with every $\lambda(g)$, and put $a=T\delta_e\in \ell^2(G)$. Since $\delta_t=\lambda(t)\delta_e$, we have \begin{align*} T\delta_t=T\lambda(t)\delta_e=\lambda(t)T\delta_e=\lambda(t)a. \end{align*} Therefore, for each $h,t\in G$, \begin{align*} (T\delta_t)(h)=(\lambda(t)a)(h)=a(t^{-1}h). \end{align*} For a finite set $F\subseteq G$, define \begin{align*} R_F=\sum_{s\in F}a(s)\rho(s^{-1}). \end{align*} Each $R_F$ lies in the algebra generated by $\rho(G)$. On a basis vector $\delta_t$, \begin{align*} (R_F\delta_t)(h)=\sum_{s\in F}a(s)(\rho(s^{-1})\delta_t)(h). \end{align*} Now $(\rho(s^{-1})\delta_t)(h)=\delta_t(hs^{-1})$, which equals $1$ exactly when $h=ts$, and equals $0$ otherwise. Hence \begin{align*} (R_F\delta_t)(h)=1_F(t^{-1}h)a(t^{-1}h). \end{align*} Comparing this with $(T\delta_t)(h)=a(t^{-1}h)$ gives \begin{align*} \|T\delta_t-R_F\delta_t\|_2^2=\sum_{h\in G}|1_{G\setminus F}(t^{-1}h)a(t^{-1}h)|^2. \end{align*} Changing variables $s=t^{-1}h$ gives \begin{align*} \|T\delta_t-R_F\delta_t\|_2^2=\sum_{s\in G\setminus F}|a(s)|^2. \end{align*} Since $a\in\ell^2(G)$, this tends to $0$ as $F$ increases through the finite subsets of $G$. Thus the finite Fourier sums built from the right translations recover $T$ on every basis vector, and the standard Fourier approximation lemma for discrete group von Neumann algebras promotes this basis-vector convergence for bounded commuting operators to strong operator convergence. Hence $T$ lies in the von Neumann algebra generated by $\rho(G)$. Therefore \begin{align*} \lambda(G)'=\rho(G)''. \end{align*} Equivalently, the commutant of the von Neumann algebra generated by the left regular representation is the von Neumann algebra generated by the right regular representation.[/proof] The vector $\delta_e$ is the bookkeeping device: translating it by $\lambda(t)$ gives every basis vector $\delta_t$, and the values of $T\delta_e$ become the convolution kernel that determines $T$ on all of $\ell^2(G)$. [/example] ## The Bicommutant Theorem The central problem of this chapter is to compare two closures of a concrete algebra: the topological closure in weak operator topology and the [algebraic closure](/page/Algebraic%20Closure) obtained by taking two commutants. The inclusion from weak closure to double commutant is direct; the content is that the double commutant contains no additional operators. [quotetheorem:1246] [citeproof:1246] This result is important because it shows that no norm approximation is being claimed. The approximating elements of $A$ depend on finitely many test vectors, which is exactly the nature of strong and weak operator convergence. The self-adjointness hypothesis is needed to make the subspace $K\subseteq H^m$ reducing; for non-self-adjoint algebras, invariant subspace projections need not commute with the algebra, so the finite-vector approximation argument breaks. The unit prevents the generated cyclic subspaces from losing the original test vectors; without a unit one usually has to pass to a unitisation before the bicommutant statement has this form. The theorem therefore identifies weak and strong operator closures for concrete unital *-algebras, but it does not identify their norm closure. [remark: Why The Unit And Adjoints Matter] The unit ensures that the generated subspaces contain the original test vectors, while closure under adjoints turns invariant subspaces into reducing subspaces. Without self-adjointness, the projection onto an invariant subspace need not commute with the algebra, and the projection argument in the proof breaks down. [/remark] The theorem gives an efficient test for being a von Neumann algebra. Instead of checking all weak operator limits, one can compute a commutant twice. [example: Diagonal Von Neumann Algebra] Let $D\subseteq \mathcal{L}(\ell^2(\mathbb N))$ be the algebra of bounded diagonal operators, and let $P_k$ denote the rank-one projection onto $\mathbb C e_k$. Each $P_k$ lies in $D$. We show that $D'=D$. Let $T\in D'$ and write $t_{ij}=(Te_j,e_i)_H$. Since $T$ commutes with $P_i$, for every $i,j$ we have \begin{align*} P_iTe_j=TP_ie_j. \end{align*} The left-hand side is \begin{align*} P_iTe_j=P_i\left(\sum_m t_{mj}e_m\right)=t_{ij}e_i. \end{align*} The right-hand side is \begin{align*} TP_ie_j=T(\delta_{ij}e_j)=\delta_{ij}Te_j. \end{align*} If $i\neq j$, then this gives $t_{ij}e_i=0$, so $t_{ij}=0$. Therefore every off-diagonal matrix coefficient of $T$ is zero, and $T$ is diagonal: $Te_j=t_{jj}e_j$ for every $j$. Since $T$ is bounded, the diagonal sequence $(t_{jj})$ is bounded, so $T\in D$. Conversely, if $T\in D$ and $S\in D$, write $Te_j=a_j e_j$ and $Se_j=b_j e_j$ with bounded scalar sequences $(a_j)$ and $(b_j)$. Then for every $j$, \begin{align*} TSe_j=T(b_j e_j)=b_j a_j e_j. \end{align*} Also \begin{align*} STe_j=S(a_j e_j)=a_j b_j e_j. \end{align*} Since $a_jb_j=b_ja_j$, we have $TSe_j=STe_j$ for every basis vector $e_j$, hence $TS=ST$ on all of $\ell^2(\mathbb N)$ by boundedness. Thus $D\subseteq D'$, and the previous paragraph gives $D'=D$. It follows that \begin{align*} D''=(D')'=D'=D. \end{align*} Thus the diagonal algebra equals its double commutant, so the criterion in *Von Neumann Bicommutant Theorem* identifies $D$ as a von Neumann algebra. [/example] The bicommutant theorem is often used alongside Kaplansky density, which refines approximation by controlling operator norms. This refinement matters when passing from a C*-algebra representation to the von Neumann algebra it generates. [quotetheorem:9266] This result is stated here in the operator-topology form used throughout the course. The norm-closed self-adjoint hypothesis is essential because the proof uses the C*-algebra structure of $A$; for a merely algebraic *-subalgebra, strong density need not preserve the unit ball. The theorem strengthens the bicommutant theorem by adding [uniform norm](/page/Uniform%20Norm) control, but it still asserts strong operator density rather than norm density. Its role in the sequel is to let approximations inside a represented C*-algebra pass to the von Neumann algebra it generates without losing bounds such as $\|a\|\le 1$. [example: Tensor Amplification] Let $H=\mathbb C^n\otimes K$ and let $A=M_n(\mathbb C)\otimes I_K$. The block matrix computation above gives \begin{align*} A'=I_n\otimes \mathcal L(K). \end{align*} We now compute the second commutant. Let $T\in A''$, and write \begin{align*} T(e_j\otimes \eta)=\sum_{i=1}^n e_i\otimes T_{ij}\eta \end{align*} with $T_{ij}\in\mathcal L(K)$. Since $T$ commutes with every $I_n\otimes R$ for $R\in\mathcal L(K)$, for every $j$ and $\eta\in K$ we have \begin{align*} T(I_n\otimes R)(e_j\otimes \eta)=\sum_{i=1}^n e_i\otimes T_{ij}R\eta. \end{align*} Also \begin{align*} (I_n\otimes R)T(e_j\otimes \eta)=\sum_{i=1}^n e_i\otimes RT_{ij}\eta. \end{align*} Comparing the $e_i$-coefficients gives \begin{align*} T_{ij}R\eta=RT_{ij}\eta \end{align*} for every $R\in\mathcal L(K)$ and every $\eta\in K$. Hence each block $T_{ij}$ commutes with all of $\mathcal L(K)$. An operator on $K$ that commutes with all of $\mathcal L(K)$ must be scalar. Indeed, if $C\in\mathcal L(K)$ commutes with every rank-one projection $P_u$ onto $\mathbb C u$, then $P_uCu=CP_uu=Cu$, so $Cu\in\mathbb C u$. Thus $Cu=\alpha_u u$ for each unit vector $u$. If $u$ and $v$ are linearly independent, applying the same fact to $u+v$ gives \begin{align*} C(u+v)=\alpha_u u+\alpha_v v=\alpha_{u+v}u+\alpha_{u+v}v, \end{align*} so $\alpha_u=\alpha_v=\alpha_{u+v}$. Therefore $C=\alpha I_K$ for one scalar $\alpha$. Thus each block has the form $T_{ij}=\alpha_{ij}I_K$, and therefore \begin{align*} T=\left(\alpha_{ij}\right)_{i,j=1}^n\otimes I_K\in M_n(\mathbb C)\otimes I_K. \end{align*} So $A''\subseteq A$. Conversely, if $B=(\beta_{ij})\otimes I_K\in A$ and $R\in\mathcal L(K)$, then for every $j$ and $\eta\in K$, \begin{align*} B(I_n\otimes R)(e_j\otimes \eta)=\sum_{i=1}^n e_i\otimes \beta_{ij}R\eta. \end{align*} Also \begin{align*} (I_n\otimes R)B(e_j\otimes \eta)=\sum_{i=1}^n e_i\otimes R(\beta_{ij}\eta)=\sum_{i=1}^n e_i\otimes \beta_{ij}R\eta. \end{align*} Hence $B$ commutes with every $I_n\otimes R$, so $A\subseteq A''$. Combining the two inclusions, \begin{align*} A''=A. \end{align*} By the *Von Neumann Bicommutant Theorem*, $A$ is weak operator closed. Thus the multiplicity space $K$ appears in the commutant $I_n\otimes\mathcal L(K)$, while the original algebra remains exactly the visible matrix action $M_n(\mathbb C)\otimes I_K$. [/example] The tensor example closes the chapter by showing the bicommutant theorem in a representation with nontrivial multiplicity space. This prepares the later study of centres and factors, where the size of the commutant measures how far a representation is from being irreducible. [explanation: Role Of The Bicommutant Theorem In The Course] The bicommutant theorem lets us pass between three languages: generators, invariant projections, and operator-topology closure. In later chapters, factors will be detected by the centre $M\cap M'$, traces will interact with commutants through standard representations, and Murray-von Neumann equivalence will be formulated using projections inside a von Neumann algebra. The present chapter supplies the mechanism that makes these later algebraic constructions compatible with weak and strong limits. [/explanation] Once commutants are in place, the internal geometry of a von Neumann algebra becomes visible through projections, partial isometries, and spectral decompositions. The same weak and strong closure mechanisms that powered the bicommutant theorem now let these geometric objects behave well under limits and support a comparison theory inside the algebra. # 3. Projections, Partial Isometries, and Internal Geometry After Chapter 2 converted weak closure into commutant language, this chapter turns the abstract closure properties of von Neumann algebras into internal geometry. It uses the preceding chapter's bicommutant theorem, the weak and strong operator topologies on $\mathcal{L}(H)$, the spectral theorem for bounded self-adjoint operators, and basic Hilbert space facts about orthogonal projections. In a Hilbert space, closed subspaces are encoded by orthogonal projections; in a von Neumann algebra, the projections belonging to the algebra record the subspaces visible to that algebra. We will see that these projections form a complete lattice, that positive operators and normal functionals have support projections, and that [polar decomposition](/theorems/3074) can be performed without leaving the algebra. ## Projection Lattices in a Von Neumann Algebra The first question is how much of Hilbert space geometry remains internal to a von Neumann algebra $M \subseteq \mathcal{L}(H)$. If two closed subspaces are represented by projections in $M$, then their intersection, closed span, and orthogonal complement should also be represented inside $M$ if $M$ is to behave like a closed algebra of observables. [definition: Projection In A Von Neumann Algebra] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra. A projection in $M$ is an operator $p:H\to H$ with $p \in M$ satisfying $p = p^* = p^2$. [/definition] A projection $p$ is the operator-theoretic representative of the closed subspace $pH$. To compare such internal subspaces, we need an order relation that can be read both algebraically inside $M$ and geometrically on $H$. [definition: Projection Order] Let $p,q \in M$ be projections. Write $p \le q$ if $pq = p$. [/definition] For projections, the relation $p \le q$ is equivalent to $pH \subseteq qH$, and also to $qp=p$. This equivalence lets us pass between operator equations and subspace geometry, and it gives the first concrete test for how the projection order behaves in familiar algebras. [example: Rank-One Projections in $B(H)$] Let $H$ be a Hilbert space and let $\xi\in H$ satisfy $\|\xi\|_H=1$. Define $p_\xi(\eta)=(\eta,\xi)_H\xi$. Then \begin{align*} p_\xi^2(\eta)=p_\xi((\eta,\xi)_H\xi)=((\eta,\xi)_H\xi,\xi)_H\xi=(\eta,\xi)_H(\xi,\xi)_H\xi=(\eta,\xi)_H\xi=p_\xi(\eta). \end{align*} Also, for $\eta,\theta\in H$, \begin{align*} (p_\xi\eta,\theta)_H=((\eta,\xi)_H\xi,\theta)_H=(\eta,\xi)_H(\xi,\theta)_H=(\eta,(\theta,\xi)_H\xi)_H=(\eta,p_\xi\theta)_H. \end{align*} Thus $p_\xi=p_\xi^*=p_\xi^2$, so $p_\xi$ is a projection. Its range is $\mathbb C\xi$, because every value of $p_\xi$ is a scalar multiple of $\xi$, and $p_\xi(\xi)=(\xi,\xi)_H\xi=\xi$. Now let $\xi,\zeta$ be unit vectors. For $\eta\in H$, \begin{align*} (p_\xi p_\zeta)(\eta)=p_\xi((\eta,\zeta)_H\zeta)=(\eta,\zeta)_H(\zeta,\xi)_H\xi. \end{align*} If $\mathbb C\xi\subseteq\mathbb C\zeta$, then $\xi=\lambda\zeta$ for some $|\lambda|=1$, so $p_\xi=p_\zeta$, and hence $p_\xi p_\zeta=p_\xi$. Conversely, if $p_\xi\le p_\zeta$, then $p_\xi p_\zeta=p_\xi$. Applying both sides to $\xi$ gives \begin{align*} (p_\xi p_\zeta)(\xi)=(\xi,\zeta)_H(\zeta,\xi)_H\xi=|(\xi,\zeta)_H|^2\xi=p_\xi(\xi)=\xi. \end{align*} Since $\xi\ne 0$, this forces $|(\xi,\zeta)_H|=1$. The equality case of the [Cauchy-Schwarz inequality](/theorems/432) for unit vectors gives $\xi\in\mathbb C\zeta$, so $\mathbb C\xi\subseteq\mathbb C\zeta$. Therefore the projection order on rank-one projections is exactly inclusion of the one-dimensional ranges, which means two such projections are comparable only when their ranges coincide. [/example] The rank-one case confirms that the order is just inclusion of visible subspaces. Once projections are ordered, the next problem is to name the projection corresponding to an intersection, to a closed span, and to an orthogonal complement when such projections exist in $M$. [definition: Meet Join And Orthocomplement] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra and let $\mathcal{P}(M)$ denote its set of projections. For a family $(p_i)_{i\in I}$ in $\mathcal{P}(M)$, a meet $\bigwedge_{i\in I}p_i$ is the greatest projection $r\in\mathcal{P}(M)$ with $r\le p_i$ for all $i\in I$. A join $\bigvee_{i\in I}p_i$ is the least projection $s\in\mathcal{P}(M)$ with $p_i\le s$ for all $i\in I$. The orthocomplement of $p\in\mathcal{P}(M)$ is $1-p$. [/definition] The definitions are order-theoretic, so they do not by themselves say that the required projections exist. The decisive closure question is whether arbitrary intersections and closed spans of ranges still come from projections in $M$. [quotetheorem:9267] [citeproof:9267] This theorem is one of the main ways weak operator topology (WOT) closure enters the internal structure of $M$. It does not say that every norm-closed $*$-subalgebra of $\mathcal{L}(H)$ has the same completeness property. For a concrete failure, let $A=c$ be the algebra of convergent scalar sequences acting diagonally on $\ell^2(\mathbb N)$, and let $p_n$ be the projection onto $\mathbb C e_{2n}$. Each $p_n$ belongs to $A$, but their join in $\mathcal{L}(\ell^2(\mathbb N))$ is the diagonal projection onto the closed span of the even basis vectors, whose diagonal sequence $1,0,1,0,\dots$ is not convergent and hence is not in $A$. Thus the theorem is a genuinely von Neumann algebra statement, not a general $C^*$-algebra statement. The next example shows what the lattice theorem becomes in a commutative von Neumann algebra where projections are characteristic functions. [example: Range Projections Of Multiplication Operators] Let $(X,\mathcal E,\mu)$ be a measure space, and let $L^\infty(X,\mu)$ act on $L^2(X,\mu)$ by $M_fh=fh$. For $E\in\mathcal E$ and $h\in L^2(X,\mu)$, \begin{align*}M_{\mathbf{1}_E}^2h=\mathbf{1}_E(\mathbf{1}_Eh)=\mathbf{1}_E^2h=\mathbf{1}_Eh=M_{\mathbf{1}_E}h.\end{align*} Also, since $\mathbf{1}_E$ is real-valued, \begin{align*}(M_{\mathbf{1}_E}h,k)_{L^2}=\int_X \mathbf{1}_Eh\overline{k}\,d\mu=\int_X h\overline{\mathbf{1}_Ek}\,d\mu=(h,M_{\mathbf{1}_E}k)_{L^2}.\end{align*} Thus $M_{\mathbf{1}_E}=M_{\mathbf{1}_E}^*=M_{\mathbf{1}_E}^2$, so it is a projection. Its range is $L^2(E)$, interpreted as the closed subspace of functions vanishing a.e. on $X\setminus E$. Indeed, $M_{\mathbf{1}_E}h=\mathbf{1}_Eh$ vanishes a.e. off $E$, while if $k$ vanishes a.e. off $E$, then $\mathbf{1}_Ek=k$ a.e., so $k=M_{\mathbf{1}_E}k$. For two measurable sets $E,F$, the product of the corresponding projections is \begin{align*}M_{\mathbf{1}_E}M_{\mathbf{1}_F}h=\mathbf{1}_E\mathbf{1}_Fh=\mathbf{1}_{E\cap F}h=M_{\mathbf{1}_{E\cap F}}h.\end{align*} Hence $M_{\mathbf{1}_{E\cap F}}\le M_{\mathbf{1}_E}$ and $M_{\mathbf{1}_{E\cap F}}\le M_{\mathbf{1}_F}$, since multiplying by $\mathbf{1}_E$ or $\mathbf{1}_F$ leaves $\mathbf{1}_{E\cap F}$ unchanged. Conversely, every projection in the multiplication algebra has the form $M_{\mathbf{1}_A}$ modulo null sets: if $M_\phi$ is a projection, then $\phi=\overline{\phi}$ and $\phi^2=\phi$ a.e., so $\phi(x)\in\{0,1\}$ a.e. If $M_{\mathbf{1}_A}\le M_{\mathbf{1}_E}$ and $M_{\mathbf{1}_A}\le M_{\mathbf{1}_F}$, then $\mathbf{1}_A\mathbf{1}_E=\mathbf{1}_A$ and $\mathbf{1}_A\mathbf{1}_F=\mathbf{1}_A$ a.e., so $A\setminus E$ and $A\setminus F$ are null, and therefore $A\setminus(E\cap F)$ is null. Thus $M_{\mathbf{1}_A}\le M_{\mathbf{1}_{E\cap F}}$, proving \begin{align*}M_{\mathbf{1}_E}\wedge M_{\mathbf{1}_F}=M_{\mathbf{1}_{E\cap F}}.\end{align*} Similarly, $M_{\mathbf{1}_E}\le M_{\mathbf{1}_{E\cup F}}$ and $M_{\mathbf{1}_F}\le M_{\mathbf{1}_{E\cup F}}$. If $M_{\mathbf{1}_E}\le M_{\mathbf{1}_A}$ and $M_{\mathbf{1}_F}\le M_{\mathbf{1}_A}$, then $E\setminus A$ and $F\setminus A$ are null, so $(E\cup F)\setminus A$ is null. Hence $M_{\mathbf{1}_{E\cup F}}\le M_{\mathbf{1}_A}$, proving \begin{align*}M_{\mathbf{1}_E}\vee M_{\mathbf{1}_F}=M_{\mathbf{1}_{E\cup F}}.\end{align*} So in this commutative von Neumann algebra, the projection lattice is exactly the measurable-set lattice modulo null sets. [/example] Projection lattices are not usually Boolean, because two projections may fail to commute. The commutative multiplication example is the special case in which the lattice reduces to measure algebra geometry, so it is important to isolate exactly where the Boolean formulas survive. [remark: Noncommutativity Of Joins And Meets] If $p,q\in\mathcal{P}(M)$ commute, then $p\wedge q=pq$ and $p\vee q=p+q-pq$. Without commutativity, $pq$ need not be a projection, so lattice operations cannot be recovered by the elementary Boolean formulas. [/remark] ## Support Projections Of Positive Operators The next question is how an element of $M$ determines the part of the Hilbert space on which it acts nondegenerately. For a positive operator, the relevant subspace is the closure of its range, or equivalently the orthogonal complement of its kernel. [definition: Support Projection Of A Positive Operator] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra and let $a:H\to H$ be a positive operator with $a\in M$. The support projection of $a$ is the projection $s(a)\in M$ onto $\overline{aH}$. [/definition] The existence of $s(a)$ inside $M$ follows from the spectral theorem: it is the spectral projection of $a$ associated to $(0,\infty)$. To use this projection in calculations, we need intrinsic characterisations that mention only $a$, the order on $M$, and projections in $M$. [quotetheorem:9268] [citeproof:9268] These identities are the operator-algebraic version of the statement that a positive function lives exactly on the set where it is nonzero. Positivity is essential for the order characterisation: if $u=-1$ on $H$, then its range support is $1$, but the inequality $u\le \|u\|_{\mathrm{op}}p$ is already true for $p=0$, so the smallest projection satisfying that inequality is not the range support. The von Neumann algebra hypothesis is also doing work. In the diagonal algebra $c\subseteq\ell^\infty$ acting on $\ell^2(\mathbb N)$, a positive diagonal operator whose entries are positive exactly on the even indices and tend to $0$ has support equal to the even-coordinate projection, which is not in $c$. Thus the theorem identifies support projections internally only under positivity and WOT closure. In a commutative von Neumann algebra, support projections become characteristic functions of essential supports, giving a useful check on the abstract definition. [example: Essential Support In A Multiplication Algebra] Let $S=\{x\in X:f(x)>0\}$, understood modulo null sets. First $M_{\mathbf{1}_S}$ is a projection, since for $h\in L^2(X,\mu)$, \begin{align*} M_{\mathbf{1}_S}^2h=\mathbf{1}_S(\mathbf{1}_Sh)=\mathbf{1}_S^2h=\mathbf{1}_Sh=M_{\mathbf{1}_S}h. \end{align*} Also $\mathbf{1}_S$ is real-valued, so $M_{\mathbf{1}_S}^*=M_{\mathbf{1}_S}$. We show that this projection is the support projection of $M_f$ by identifying $\overline{M_fL^2(X,\mu)}$. For every $h\in L^2(X,\mu)$, the function $fh$ vanishes a.e. on $X\setminus S$, so $M_fL^2(X,\mu)\subseteq L^2(S)$. Conversely, let $k\in L^2(S)$ and define \begin{align*} h_n=\mathbf{1}_{\{f\ge 1/n\}}\frac{k}{f}. \end{align*} On $\{f\ge 1/n\}$ we have $|1/f|\le n$, so $|h_n|\le n|k|$ and hence $h_n\in L^2(X,\mu)$. Then \begin{align*} M_fh_n=f\mathbf{1}_{\{f\ge 1/n\}}\frac{k}{f}=\mathbf{1}_{\{f\ge 1/n\}}k. \end{align*} The sets $\{f\ge 1/n\}$ increase to $S$ modulo null sets, and \begin{align*} |k-\mathbf{1}_{\{f\ge 1/n\}}k|^2\le |k|^2. \end{align*} By dominated convergence, \begin{align*} \|k-M_fh_n\|_2^2=\int_X |k-\mathbf{1}_{\{f\ge 1/n\}}k|^2\,d\mu\to 0. \end{align*} Thus $L^2(S)\subseteq\overline{M_fL^2(X,\mu)}$, so $\overline{M_fL^2(X,\mu)}=L^2(S)$ and therefore $s(M_f)=M_{\mathbf{1}_S}$. Finally, the operator inequality $M_f\le \|f\|_\infty M_{\mathbf{1}_S}$ follows from the pointwise inequality $0\le f\le \|f\|_\infty\mathbf{1}_S$ a.e. Indeed, for $h\in L^2(X,\mu)$, \begin{align*} ((\|f\|_\infty M_{\mathbf{1}_S}-M_f)h,h)_{L^2}=\int_X(\|f\|_\infty\mathbf{1}_S-f)|h|^2\,d\mu\ge 0. \end{align*} So the abstract support projection is exactly multiplication by the characteristic function of the essential set on which $f$ is nonzero. [/example] Normal functionals also have supports, but now the projection records where the functional sees the algebra. For positive normal functionals this is a noncommutative analogue of the support of a measure, and the complete lattice theorem gives a way to define it by joining all projections on which the functional vanishes. [definition: Support Projection Of A Positive Normal Functional] Let $M$ be a von Neumann algebra and let $\varphi:M\to\mathbb C$ be a positive normal functional, so $\varphi\in M_*^+$. The support projection of $\varphi$ is \begin{align*} s(\varphi):=1-\bigvee\{p\in\mathcal P(M):\varphi(p)=0\}. \end{align*} [/definition] The projection $1-s(\varphi)$ is the largest projection ignored by $\varphi$. Thus $s(\varphi)$ is intended to be the smallest part of the identity on which $\varphi$ can be faithful. The remaining issue is that the definition only mentions projections, while applications must know what happens to arbitrary elements whose square $x^*x$ is invisible to $\varphi$. The following result bridges that gap between projection support and elementwise faithfulness. [quotetheorem:9269] [citeproof:9269] This theorem is the bridge between order, normality, and Hilbert-space-like orthogonality. Positivity is needed because the condition $\varphi(x^*x)=0$ uses the positive cone to behave like a squared norm. A concrete failure occurs already in the finite-dimensional von Neumann algebra $M=\mathbb C\oplus\mathbb C$: the normal functional $\varphi(\lambda,\mu)=\lambda-\mu$ is not positive, and for $x=(1,1)$ we have $\varphi(x^*x)=\varphi(1,1)=0$ although $x$ is nonzero on both central summands. The projection formula based on annihilated projections also loses its intended meaning here, since $\varphi(1)=0$ even though $\varphi$ is not the zero functional. Normality is needed at the join step in the proof. For example, on $M=\ell^\infty(\mathbb N)$ there are singular states extending the limit functional on convergent sequences. Such a state $\psi$ satisfies $\psi(e_n)=0$ for each singleton projection $e_n$, while $\bigvee_{n=1}^{\infty}e_n=1$ and $\psi(1)=1$, so vanishing is not preserved under this supremum. The theorem therefore does not assign a support projection with the same faithfulness property to arbitrary bounded functionals on $M$. It will later be used to compare projections via states and traces, and vector functionals provide the most concrete noncommutative example. [example: Vector Functional Support] Let $M\subseteq\mathcal{L}(H)$ be a von Neumann algebra, fix $\xi\in H$, and define $\omega_\xi(x)=(x\xi,\xi)_H$. Put $K=\overline{M'\xi}$, and let $e$ be the orthogonal projection onto $K$. Since $M'\xi\subseteq K$, the subspace $K$ is invariant under every $y\in M'$, and it is also invariant under $y^*$ because $y^*\in M'$. Hence $K$ reduces $M'$, so $e$ commutes with every element of $M'$. By the *Bicommutant Theorem*, $e\in M$. We show that $s(\omega_\xi)=e$. Let $p\in\mathcal P(M)$. Since $p=p^*=p^2$, \begin{align*} \omega_\xi(p)=(p\xi,\xi)_H=(p\xi,p\xi)_H+(p\xi,(1-p)\xi)_H=\|p\xi\|_H^2+((1-p)p\xi,\xi)_H=\|p\xi\|_H^2. \end{align*} Thus $\omega_\xi(p)=0$ if and only if $p\xi=0$. Now $q:=1-e$ satisfies $q\xi=0$, because $\xi=1\xi\in M'\xi\subseteq K$, so $\omega_\xi(q)=0$. Conversely, if $p\in\mathcal P(M)$ and $\omega_\xi(p)=0$, then $p\xi=0$. For $\eta\in pH$ and $y\in M'$, \begin{align*} (\eta,y\xi)_H=(p\eta,y\xi)_H=(\eta,py\xi)_H=(\eta,yp\xi)_H=0. \end{align*} Therefore $pH\perp M'\xi$, so $pH\subseteq K^\perp=qH$, which means $p\le q$. Hence $q$ is the largest projection annihilated by $\omega_\xi$, and the definition of support gives \begin{align*} s(\omega_\xi)=1-q=e. \end{align*} So the support of the vector functional is exactly the projection in $M$ onto $\overline{M'\xi}$; it records the cyclic subspace generated from $\xi$ by the commutant. [/example] ## Polar Decomposition Inside A Von Neumann Algebra The final question in this chapter is whether an arbitrary element of $M$ has its geometric factorisation inside $M$. On a Hilbert space, every bounded operator splits into a partial isometry and a positive operator; for von Neumann algebras, the point is that both factors remain internal. [definition: Partial Isometry] Let $H$ be a Hilbert space. An operator $v:H\to H$ with $v\in\mathcal{L}(H)$ is a partial isometry if $v^*v$ is a projection. The projection $v^*v$ is the initial projection of $v$, and $vv^*$ is the final projection of $v$. [/definition] A partial isometry is unitary from its initial subspace onto its final subspace and is zero on the orthogonal complement of the initial subspace. Thus partial isometries are the correct morphisms between projections, as the basic closed-subspace construction shows. [example: Partial Isometry Between Closed Subspaces] Let $K,L\subseteq H$ be closed subspaces, let $u:K\to L$ be unitary, and write each $\eta\in H$ uniquely as $\eta=\eta_K+\eta_{K^\perp}$ with $\eta_K\in K$ and $\eta_{K^\perp}\in K^\perp$. Define $v\eta=u\eta_K$. Then \begin{align*} \|v\eta\|_H^2=\|u\eta_K\|_H^2=\|\eta_K\|_H^2\le \|\eta_K\|_H^2+\|\eta_{K^\perp}\|_H^2=\|\eta\|_H^2. \end{align*} Thus $v$ is bounded. If $\theta=\theta_L+\theta_{L^\perp}$ is the orthogonal decomposition of $\theta$ relative to $L$, then for every $\eta\in H$, \begin{align*} (v\eta,\theta)_H=(u\eta_K,\theta_L+\theta_{L^\perp})_H=(u\eta_K,\theta_L)_H. \end{align*} Since $u:K\to L$ is unitary, \begin{align*} (u\eta_K,\theta_L)_H=(\eta_K,u^{-1}\theta_L)_H. \end{align*} Because $u^{-1}\theta_L\in K$ and $\eta_{K^\perp}\perp K$, \begin{align*} (\eta_K,u^{-1}\theta_L)_H=(\eta_K+\eta_{K^\perp},u^{-1}\theta_L)_H=(\eta,u^{-1}\theta_L)_H. \end{align*} Hence $v^*\theta=u^{-1}\theta_L$. Therefore \begin{align*} v^*v\eta=v^*(u\eta_K)=u^{-1}(u\eta_K)=\eta_K. \end{align*} So $v^*v$ is the orthogonal projection onto $K$. Similarly, \begin{align*} vv^*\theta=v(u^{-1}\theta_L)=u(u^{-1}\theta_L)=\theta_L. \end{align*} Thus $vv^*$ is the orthogonal projection onto $L$. Since $v^*v$ is a projection, $v$ is a partial isometry; it acts as the unitary $u$ from $K$ onto $L$ and vanishes on $K^\perp$. [/example] For an operator $x$, the positive part $|x|=(x^*x)^{1/2}$ records size, while the partial isometry records the direction from the support of $|x|$ to the support of $xx^*$. The key internal question is whether this direction operator belongs to $M$ whenever $x$ does. [quotetheorem:9270] [citeproof:9270] The theorem says that $M$ contains enough partial isometries to compare the ranges of its own elements. It does not say that polar decomposition stays inside an arbitrary norm-closed $*$-algebra. For a concrete commutative failure, let $A=C([0,1])$ act on $L^2([0,1])$ by multiplication and take $g(t)=\max(t-1/2,0)$. The polar partial isometry of $M_g$ is multiplication by $\mathbf{1}_{(1/2,1]}$, which is not continuous and therefore is not an element of $A$. WOT closure is what replaces this missing characteristic function by an internal projection in the von Neumann algebra case. The theorem also does not make the polar partial isometry unitary unless the source and range supports are both $1$. This is the local geometry that later becomes Murray-von Neumann equivalence of projections, and the commutative case reduces to the usual phase of a function. [example: Polar Decomposition Of A Multiplication Operator] Let $g\in L^\infty(X,\mu)$, let $M_g$ act on $L^2(X,\mu)$, and put $S=\{x\in X:g(x)\ne 0\}$ modulo null sets. Define $u(x)=g(x)/|g(x)|$ for $x\in S$ and $u(x)=0$ for $x\notin S$. Then $|u(x)|=1$ on $S$ and $|u(x)|=0$ off $S$, so $u\in L^\infty(X,\mu)$ and $|u|^2=\mathbf{1}_S$ a.e. For $h\in L^2(X,\mu)$ and a.e. $x\in X$, we have \begin{align*}(M_uM_{|g|}h)(x)=u(x)|g(x)|h(x).\end{align*} If $x\in S$, then $u(x)|g(x)|=g(x)$ by the definition of $u$; if $x\notin S$, then $g(x)=0$ and $u(x)|g(x)|=0$. Hence \begin{align*}u|g|h=gh.\end{align*} Therefore \begin{align*}M_uM_{|g|}=M_g.\end{align*} Next, for $h,k\in L^2(X,\mu)$, \begin{align*}(M_uh,k)_{L^2}=\int_X uh\overline{k}\,d\mu=\int_X h\overline{\overline{u}k}\,d\mu=(h,M_{\overline{u}}k)_{L^2}.\end{align*} Thus $M_u^*=M_{\overline{u}}$. Hence, for every $h\in L^2(X,\mu)$, \begin{align*}(M_u^*M_uh)(x)=\overline{u(x)}u(x)h(x)=|u(x)|^2h(x)=\mathbf{1}_S(x)h(x).\end{align*} So \begin{align*}M_u^*M_u=M_{\mathbf{1}_S}.\end{align*} Similarly, \begin{align*}(M_uM_u^*h)(x)=u(x)\overline{u(x)}h(x)=|u(x)|^2h(x)=\mathbf{1}_S(x)h(x),\end{align*} and therefore \begin{align*}M_uM_u^*=M_{\mathbf{1}_S}.\end{align*} Since $M_{\mathbf{1}_S}$ is a projection, $M_u$ is a partial isometry. Thus the polar decomposition of $M_g$ is \begin{align*}M_g=M_uM_{|g|},\end{align*} with both initial and final projection equal to $M_{\mathbf{1}_{\{g\ne 0\}}}$. In the commutative multiplication algebra, the polar partial isometry is just multiplication by the phase of $g$, extended by $0$ on the zero set of $g$. [/example] The multiplication example is commutative, so the initial and final projections agree. In a noncommutative algebra, the two support projections need not be equal, and the partial isometry keeps track of how $x$ transports one projection onto the other. [remark: Source And Range Supports] For $x\in M$, the projection $s(|x|)=s(x^*x)$ is the source support of $x$, while $s(xx^*)$ is the range support of $x$. The polar partial isometry $v$ satisfies $v^*v=s(x^*x)$ and $vv^*=s(xx^*)$, so it implements an isometric identification between these two projections. [/remark] These three themes fit together: complete projection lattices provide the internal subspaces, support projections attach canonical projections to operators and functionals, and polar decomposition supplies the partial isometries connecting those projections. The following chapters use this geometry to compare projections, define equivalence, and begin the type classification of von Neumann algebras. Projections and partial isometries describe the internal geometry of a von Neumann algebra, but the next chapter shifts from geometry to the functionals that respect it. Normal representations and the predual encode the weak-* structure that makes ultraweak continuity, monotone convergence, and state spaces fit together naturally. # 4. Normal Representations and Preduals This chapter turns the representation theory developed earlier into a language for maps and functionals that respect the order structure of a von Neumann algebra. In a $C^*$-algebra, norm-continuity is built into every bounded linear functional, but von Neumann algebras carry a finer topology coming from their concrete action on Hilbert space. The central question is which representations, states, and positive maps preserve the suprema of projections and increasing bounded nets. The answer is normality, and the resulting space of normal functionals is the predual $M_*$. This is the noncommutative analogue of the passage from $L^\infty$ to its countably additive $L^1$ predual, and it is also the topology behind density matrices in quantum statistical mechanics. ## Normal Representations and Suprema of Projections A representation of a $C^*$-algebra need not remember the weak operator closure that makes a von Neumann algebra special. The concrete failure is already visible in commutative examples: a state on $\ell^\infty(\mathbb N)$ coming from a free ultrafilter sends the increasing projections $\mathbb{1}_{\{1,\dots,k\}}$ to $0$ while their join is $1$. A representation or functional with this behaviour cannot support monotone convergence for spectral projections, so it loses the measure-theoretic information carried by the von Neumann algebra. For von Neumann algebras, the useful representations are therefore those compatible with increasing joins of projections, because projections encode closed subspaces, spectral decompositions, and order limits. [definition: Normal Star Homomorphism] Let $M$ and $N$ be von Neumann algebras. A unital $*$-homomorphism $\rho: M \to N$ is normal if for every bounded increasing net $(x_i)_{i \in I}$ in $M_+$ with supremum $x \in M_+$, the net $(\rho(x_i))_{i \in I}$ has supremum $\rho(x)$ in $N_+$. [/definition] This definition is order-theoretic, but in practice it can be checked on projections. That reduction is powerful because a self-adjoint element is reconstructed from its spectral projections, so preservation of joins of projections forces compatibility with the spectral calculus. [quotetheorem:9271] [citeproof:9271] The theorem says that normality is not an additional algebraic law; it is continuity with respect to the lattice of projections. The hypothesis that $\rho$ is a $*$-homomorphism matters: a merely positive unital map can preserve order without preserving spectral projections, so projection joins no longer determine its behaviour on the functional calculus. The result also does not say that every representation of the underlying $C^*$-algebra is normal; free-ultrafilter states on $\ell^\infty(\mathbb N)$ give non-normal cyclic representations. The next example fixes the concrete Hilbert space model that should be kept in mind throughout the chapter. [example: Amplification By A Multiplicity Space] Let $H=K\otimes L$ with $L\ne\{0\}$, and define $\rho:B(K)\to B(H)$ by \begin{align*} \rho(a)=a\otimes I_L. \end{align*} For $a,b\in B(K)$ and $\xi\otimes \eta\in K\otimes L$, \begin{align*} \rho(ab)(\xi\otimes \eta)=ab\xi\otimes \eta=(a\otimes I_L)(b\xi\otimes \eta)=\rho(a)\rho(b)(\xi\otimes \eta). \end{align*} Since finite sums of elementary tensors are dense, $\rho(ab)=\rho(a)\rho(b)$. Also $\rho(I_K)=I_K\otimes I_L=I_H$, and for elementary tensors, \begin{align*} ((a\otimes I_L)(\xi\otimes \eta),\xi'\otimes \eta')=(a\xi,\xi')_K(\eta,\eta')_L=(\xi,a^*\xi')_K(\eta,\eta')_L=(\xi\otimes \eta,(a^*\otimes I_L)(\xi'\otimes \eta')). \end{align*} Thus $\rho(a)^*=\rho(a^*)$, so $\rho$ is a unital $*$-homomorphism. Let $(q_i)$ be an increasing net of projections on $K$ with supremum $q$. Since $q_i\le q$, every vector in $(1-q)K$ is killed by all $q_i$. If $\xi\in qK$, then the closed span of $\bigcup_i q_iK$ is $qK$, so for every $\varepsilon>0$ there are $j$ and $\xi_0\in q_jK$ with $\|\xi-\xi_0\|<\varepsilon$. For $i\ge j$, \begin{align*} \|q_i\xi-\xi\|\le \|q_i(\xi-\xi_0)\|+\|q_i\xi_0-\xi_0\|+\|\xi_0-\xi\|<2\varepsilon. \end{align*} Hence $q_i\xi\to q\xi$ for every $\xi\in K$. For a finite tensor sum $v=\sum_{m=1}^r \xi_m\otimes \eta_m$, \begin{align*} (q_i\otimes I_L)v-(q\otimes I_L)v=\sum_{m=1}^r (q_i\xi_m-q\xi_m)\otimes \eta_m, \end{align*} so this norm tends to $0$ because each $q_i\xi_m\to q\xi_m$. For arbitrary $\zeta\in H$, choose such a finite tensor sum $v$ with $\|\zeta-v\|<\varepsilon$. Since $\|q_i\otimes I_L\|\le 1$ and $\|q\otimes I_L\|\le 1$, \begin{align*} \|(q_i\otimes I_L)\zeta-(q\otimes I_L)\zeta\|\le 2\varepsilon+\|(q_i\otimes I_L)v-(q\otimes I_L)v\|. \end{align*} Thus $q_i\otimes I_L$ converges strongly to $q\otimes I_L$. The projections $q_i\otimes I_L$ are increasing because $q_iq_j=q_i$ whenever $i\le j$, so their supremum is $q\otimes I_L$. Therefore $\rho$ preserves increasing joins of projections, and by *Projection Criterion For Normality* it is normal. This is the standard multiplicity-space representation of $B(K)$. [/example] This example shows how normality behaves for concrete Hilbert space representations. It also avoids a tempting but incorrect construction: $a\mapsto a\oplus I_L$ is not linear unless the scalar part is treated through a genuine representation of the same algebra. To name the representations allowed in the rest of the course, we now impose normality as part of the representation data. [definition: Normal Representation] Let $M$ be a von Neumann algebra and let $H$ be a Hilbert space. A representation of $M$ on $H$ is a unital normal $*$-homomorphism $\pi:M\to B(H)$. [/definition] Normal representations are the representations used in the standard form and in direct integral theory. Non-normal representations can exist, but they may collapse increasing joins of projections and therefore lose the measure-theoretic part of the algebra. ## The Predual And Normal Functionals The [dual space](/page/Dual%20Space) $M^*$ of a von Neumann algebra is too large for most operator-algebraic arguments. On $\ell^\infty(\mathbb N)$, bounded linear functionals include finitely additive probability measures on $\mathbb N$, and these do not have to be countably additive on disjoint unions. Such functionals are norm-continuous but fail monotone convergence, so norm duality alone cannot distinguish genuine measure-theoretic states from finitely additive ones. The functionals that interact well with monotone convergence form a distinguished closed subspace, and the theorem of Sakai says that this subspace is not extra structure: it characterises von Neumann algebras among $C^*$-algebras. [definition: Normal Positive Functional] Let $M$ be a von Neumann algebra. A positive functional $\varphi:M\to \mathbb C$ is normal if for every bounded increasing net $(x_i)_{i\in I}$ in $M_+$ with supremum $x\in M_+$, \begin{align*} \varphi(x)=\sup_{i\in I}\varphi(x_i). \end{align*} [/definition] The definition is the scalar analogue of normality for homomorphisms. Because positive functionals are determined by their values on the positive cone, this is the correct monotone convergence property for states and weights. [quotetheorem:9272] [citeproof:9272] The criteria explain why normal states are the states compatible with countable or directed measurement refinement, and each hypothesis is doing real work. Positivity is needed because the theorem reduces a functional to a finite measure on the projection lattice; for a non-positive functional such as $i\tau$ on a finite von Neumann algebra with trace $\tau$, the values on projections are not ordered [real numbers](/page/Real%20Numbers), so the displayed suprema are not meaningful. Normality is not automatic for positive norm-continuous functionals: the free-ultrafilter state on $\ell^\infty(\mathbb N)$ sends the projections $\mathbb{1}_{\{1,\dots,k\}}$ to $0$ although their supremum is $1$. Orthogonality is also essential in the additive condition. For overlapping projections $p$ and $q$, additivity would demand $\varphi(p\vee q)=\varphi(p)+\varphi(q)$ even when $p=q=1$, which is incompatible with any state because it would give $1=2$. Thus the theorem separates genuinely countably or completely additive behaviour from singular finitely additive behaviour. This separation motivates treating the normal part of $M^*$ as its own [Banach space](/page/Banach%20Space). [definition: Predual] Let $M$ be a von Neumann algebra. A Banach space $E$ is a predual of $M$ if there is an isometric isomorphism $E^*\cong M$. [/definition] For a general dual Banach space, a predual need not be unique, and choosing the wrong one changes the [weak* topology](/page/Weak*%20Topology). That is a serious obstruction here: the topology used for compactness and continuity should be the one detected by normal states, not by all finitely additive states in $M^*$. The remarkable fact for von Neumann algebras is that the normal functionals form the canonical predual. [quotetheorem:9273] [citeproof:9273] This theorem justifies writing $M_*$ without specifying a representation. Its force is stronger than the corresponding statement for arbitrary dual Banach spaces: the operator-algebra structure rules out competing isometric preduals. The theorem also does not say that $M^*=M_*$: singular functionals still live in the full Banach dual. To discuss physical and probabilistic interpretations, we next single out the positive norm-one elements of this canonical predual. [definition: Normal State] Let $M$ be a von Neumann algebra. A normal state on $M$ is a linear functional $\varphi:M\to\mathbb C$ such that $\varphi$ is positive, normal, and satisfies $\varphi(1)=1$. [/definition] Normal states are the states that arise from density operators in the basic example $B(H)$. The problem is to identify which operators represent normal functionals on the full algebra of bounded operators: bounded operators themselves are too large, while finite-rank formulas suggest that summable singular values are the right size condition. The trace-class duality theorem makes this representation precise. [quotetheorem:9274] [citeproof:9274] The notation $\xi\otimes\eta$ denotes the rank-one operator $\zeta\mapsto (\zeta,\eta)_H\xi$. The trace-class hypothesis is necessary: compact operators that are not trace-class do not define bounded normal functionals on all of $B(H)$. The theorem also does not identify the full Banach dual of $B(H)$; singular functionals remain outside the trace-class predual. In particular, a unit vector $\xi\in H$ gives the vector state $\omega_\xi(a)=(a\xi,\xi)_H$, corresponding to the rank-one projection $\xi\otimes\xi$. [example: Density Matrices On A Finite Dimensional Hilbert Space] Let $H=\mathbb C^n$. Then $B(H)=M_n(\mathbb C)$, and every linear functional on $M_n(\mathbb C)$ is normal because the space is finite-dimensional. By *Trace Class Predual Of B H*, a normal state has the form $\omega_T(a)=\operatorname{Tr}(Ta)$ with $T\in M_n(\mathbb C)$, $T\ge 0$, and $\operatorname{Tr}(T)=1$. Since $T\ge 0$, the finite-dimensional spectral theorem gives an orthonormal basis $e_1,\dots,e_n$ and numbers $\lambda_1,\dots,\lambda_n\ge 0$ such that $Te_i=\lambda_i e_i$ for each $i$. The trace condition is therefore \begin{align*} 1=\operatorname{Tr}(T)=\sum_{i=1}^n (Te_i,e_i)_H \end{align*} and, because $Te_i=\lambda_i e_i$ and $\|e_i\|=1$, \begin{align*} \sum_{i=1}^n (Te_i,e_i)_H=\sum_{i=1}^n (\lambda_i e_i,e_i)_H=\sum_{i=1}^n \lambda_i. \end{align*} For any $a\in M_n(\mathbb C)$, the trace of $Ta$ in this basis is \begin{align*} \operatorname{Tr}(Ta)=\sum_{i=1}^n (Ta e_i,e_i)_H. \end{align*} Using cyclicity of the matrix trace in finite dimension, $\operatorname{Tr}(Ta)=\operatorname{Tr}(aT)$, so \begin{align*} \operatorname{Tr}(Ta)=\sum_{i=1}^n (aT e_i,e_i)_H. \end{align*} Substituting $Te_i=\lambda_i e_i$ gives \begin{align*} \operatorname{Tr}(Ta)=\sum_{i=1}^n (a(\lambda_i e_i),e_i)_H=\sum_{i=1}^n \lambda_i(ae_i,e_i)_H. \end{align*} Thus $\omega_T=\sum_{i=1}^n \lambda_i\omega_{e_i}$, where $\omega_{e_i}(a)=(ae_i,e_i)_H$, and the coefficients satisfy $\lambda_i\ge 0$ and $\sum_i\lambda_i=1$. A finite-dimensional density matrix is therefore exactly a convex combination of vector states, with weights given by its eigenvalues. [/example] Commutative von Neumann algebras give the measure-theoretic version of the same story. Normal states are represented by integrable densities rather than by arbitrary finitely additive measures. [example: Normal States on $L^\infty$] Let $(X,\mathcal A,\mu)$ be a $\sigma$-finite measure space, and let $M=L^\infty(X,\mathcal A,\mu)$ act on $L^2(X,\mathcal A,\mu)$ by multiplication. If $g\in L^1(X,\mathcal A,\mu)$ satisfies $g\ge 0$ and $\int_X g\,d\mu=1$, define \begin{align*} \varphi_g(f)=\int_X f g\,d\mu. \end{align*} For $f\ge 0$ in $L^\infty$, the product $fg$ is nonnegative and integrable, so $\varphi_g(f)\ge 0$. Also \begin{align*} \varphi_g(1)=\int_X g\,d\mu=1. \end{align*} Thus $\varphi_g$ is a state. To check normality, let $(f_i)_{i\in I}$ be a bounded increasing net in $L^\infty_+$ with supremum $f$. Then $f_i g\le f_j g\le fg$ whenever $i\le j$, because $g\ge 0$. Hence the integrals form an increasing bounded net: \begin{align*} \varphi_g(f_i)=\int_X f_i g\,d\mu\le \int_X f g\,d\mu=\varphi_g(f). \end{align*} The pointwise supremum of the functions $f_i g$ is $fg$ up to null sets, and the directed form of the monotone convergence theorem gives \begin{align*} \sup_i \varphi_g(f_i)=\sup_i\int_X f_i g\,d\mu=\int_X fg\,d\mu=\varphi_g(f). \end{align*} Therefore $\varphi_g$ is normal. Conversely, let $\varphi$ be a normal state on $L^\infty(X,\mathcal A,\mu)$, and define a set function \begin{align*} \nu(E)=\varphi(\mathbf 1_E). \end{align*} If $E_1,E_2,\dots$ are pairwise disjoint and $F_n=\bigcup_{k=1}^n E_k$, then \begin{align*} \mathbf 1_{F_n}=\sum_{k=1}^n \mathbf 1_{E_k}. \end{align*} Linearity gives \begin{align*} \nu(F_n)=\sum_{k=1}^n \nu(E_k). \end{align*} Since $\mathbf 1_{F_n}\uparrow \mathbf 1_{\bigcup_{k=1}^\infty E_k}$, normality gives \begin{align*} \nu\left(\bigcup_{k=1}^\infty E_k\right)=\sup_n \nu(F_n)=\sum_{k=1}^\infty \nu(E_k). \end{align*} Thus $\nu$ is a countably additive probability measure. If $\mu(E)=0$, then $\mathbf 1_E=0$ in $L^\infty$, so \begin{align*} \nu(E)=\varphi(\mathbf 1_E)=\varphi(0)=0. \end{align*} Hence $\nu\ll\mu$. By the [Radon-Nikodym theorem](/theorems/1247), there is $g\in L^1(X,\mathcal A,\mu)$ with $g\ge 0$ such that \begin{align*} \nu(E)=\int_E g\,d\mu \end{align*} for every $E\in\mathcal A$. Since $\nu(X)=\varphi(1)=1$, this $g$ satisfies \begin{align*} \int_X g\,d\mu=1. \end{align*} For a nonnegative simple function $s=\sum_{k=1}^m c_k\mathbf 1_{E_k}$ with $c_k\ge 0$, linearity gives \begin{align*} \varphi(s)=\sum_{k=1}^m c_k\varphi(\mathbf 1_{E_k})=\sum_{k=1}^m c_k\nu(E_k)=\int_X sg\,d\mu. \end{align*} Every positive $f\in L^\infty$ is the supremum of an increasing sequence of positive simple functions $s_n\le f$, so normality and monotone convergence give \begin{align*} \varphi(f)=\sup_n\varphi(s_n)=\sup_n\int_X s_n g\,d\mu=\int_X fg\,d\mu. \end{align*} Writing an arbitrary $f\in L^\infty$ as a linear combination of four positive functions extends the same formula to all $f$. Thus every normal state has the form $\varphi_g(f)=\int_X fg\,d\mu$ with $g\in L^1_+$ and $\int_X g\,d\mu=1$. In the standard $\sigma$-finite setting, the predual of $L^\infty(X,\mathcal A,\mu)$ is therefore $L^1(X,\mathcal A,\mu)$ under the pairing $g\mapsto(f\mapsto\int_X fg\,d\mu)$. For general measure spaces, the same statement is formulated on the localizable measure algebra: normal functionals are precisely the order-continuous measures absolutely continuous with respect to the measure class, represented by $L^1$ on that measure algebra. [/example] The warning example is that positivity and norm-continuity do not imply normality. The missing property is countable or directed additivity over joins of projections. [example: Non-Normal Pure States on $\ell^\infty(\mathbb N)$] Identify $\ell^\infty(\mathbb N)$ with $C(\beta\mathbb N)$, where $\beta\mathbb N$ is the Stone-Cech compactification. If $x\in\beta\mathbb N\setminus\mathbb N$, evaluation at $x$ is a character on $C(\beta\mathbb N)$, hence a pure state. Equivalently, $x$ corresponds to a free ultrafilter $\mathcal U$ on $\mathbb N$, and the state is \begin{align*} \varphi_{\mathcal U}((a_n))=\lim_{\mathcal U} a_n. \end{align*} For each $k\in\mathbb N$, let $p_k=\mathbf 1_{\{1,\dots,k\}}\in\ell^\infty(\mathbb N)$. If $k\le \ell$, then $\{1,\dots,k\}\subseteq\{1,\dots,\ell\}$, so \begin{align*} p_k(n)\le p_\ell(n)\quad\text{for every }n\in\mathbb N. \end{align*} Thus $(p_k)$ is an increasing sequence of projections. For each fixed $n\in\mathbb N$, one has $p_k(n)=1$ whenever $k\ge n$, so \begin{align*} \sup_{k\in\mathbb N}p_k(n)=1. \end{align*} Therefore the supremum of the increasing sequence $(p_k)$ in $\ell^\infty(\mathbb N)_+$ is the constant function $1$. Now fix $k$. Since $\mathcal U$ is free, it contains no finite subset of $\mathbb N$, so $\{1,\dots,k\}\notin\mathcal U$. Because $\mathcal U$ is an ultrafilter, the complement $\{k+1,k+2,\dots\}$ belongs to $\mathcal U$. On this complement, $p_k$ is identically $0$, hence \begin{align*} \varphi_{\mathcal U}(p_k)=\lim_{\mathcal U}p_k=0. \end{align*} For the constant function $1$, the ultrafilter limit is \begin{align*} \varphi_{\mathcal U}(1)=\lim_{\mathcal U}1=1. \end{align*} Consequently, \begin{align*} \sup_k\varphi_{\mathcal U}(p_k)=0\ne 1=\varphi_{\mathcal U}\left(\sup_k p_k\right). \end{align*} The state $\varphi_{\mathcal U}$ therefore fails monotone convergence for this increasing sequence of projections, so it is not normal. [/example] ## Ultraweak Continuity And Normal Positive Maps Once $M_*$ is known to be canonical, the topology it induces becomes the natural weak topology on $M$. The obstruction is practical as well as conceptual: checking a positive map on every bounded increasing net is often hard, especially for maps built from formulas or limits. A topological criterion lets us test normality by composing with normal functionals and checking continuity on bounded sets. Maps between von Neumann algebras are normal exactly when they are continuous for this topology on bounded sets, and this gives a usable criterion for positive maps and conditional expectations. [definition: Ultraweak Topology] Let $M$ be a von Neumann algebra with predual $M_*$. The ultraweak topology on $M$ is the weak* topology $\sigma(M,M_*)$ generated by the seminorms \begin{align*} x\mapsto |\omega(x)|, \qquad \omega\in M_*. \end{align*} [/definition] This topology agrees with the weak operator topology on bounded subsets after representing $M$ faithfully and normally, but it is intrinsic because it only uses $M_*$. The bounded-set qualification matters: weak and ultraweak topologies can behave differently on unbounded sets, and normality is an order-continuity property for bounded increasing nets. Normality can now be phrased as continuity rather than as preservation of suprema. [quotetheorem:9275] [citeproof:9275] The theorem applies in particular to completely positive maps, conditional expectations, and representations. Positivity is part of the statement because monotone convergence is an order condition; for arbitrary linear maps, predual-continuity is the right hypothesis instead. More precisely, a bounded linear map $\Phi:M\to N$ is ultraweakly continuous on bounded sets if and only if its Banach adjoint satisfies $\Phi^*(N_*)\subseteq M_*$. Without this topological hypothesis, a linear formula can fail normality even when it is bounded: if $\psi$ is a singular state on $M$ and $0\ne y\in N$, then $\Phi(x)=\psi(x)y$ is a bounded linear map, and it is normal exactly when $\psi$ is normal. Thus the criterion is not a substitute for complete positivity or bimodularity in applications; it only identifies when the map respects the predual topology. The following definition records the projection-like maps that preserve a subalgebra while respecting positivity. [definition: Conditional Expectation] Let $N\subseteq M$ be von Neumann algebras with the same identity. A [conditional expectation](/page/Conditional%20Expectation) from $M$ onto $N$ is a positive linear map $E:M\to N$ such that $E(1)=1$, $E(n)=n$ for all $n\in N$, and \begin{align*} E(n_1 x n_2)=n_1E(x)n_2 \end{align*} for all $x\in M$ and all $n_1,n_2\in N$. [/definition] The bimodule identity says that $E$ is not merely a Banach-space projection; it is compatible with multiplication by the subalgebra. Without that identity, a positive norm-one projection onto a linear copy of $N$ need not preserve the [conditional probability](/page/Conditional%20Probability) interpretation, because multiplying before or after conditioning could change the answer. We next record what normality adds: the expectation becomes a projection that is continuous for the canonical predual topology. [quotetheorem:9276] [citeproof:9276] Normality is the hypothesis that distinguishes this statement from purely Banach-space projection theory. A positive unital projection can fail to be ultraweakly continuous if it is built using a singular state, for instance by conditioning $\ell^\infty(\mathbb N)$ through a free-ultrafilter functional. The theorem also does not assert that a normal conditional expectation exists for every inclusion $N\subseteq M$; existence can require additional structure such as an invariant faithful normal state or a finite-index condition. The norm-one conclusion prepares the use of Russo-Dye for controlling positive maps from their action on unitaries. Norm control for positive maps is often reduced to the unitary elements of a $C^*$-algebra, but this reduction is not automatic: a general element of the unit ball need not be unitary, and estimates on unitaries would be useless unless unitaries generate the unit ball by a limiting process. The missing geometric fact is that the unit ball of a unital $C^*$-algebra is controlled by convex combinations of unitaries, which is exactly the reduction principle used repeatedly for completely positive maps and conditional expectations. [quotetheorem:9277] In this chapter the theorem is used as a norm-control device: estimates proved first on unitaries extend by convexity and norm closure to the full unit ball. The unital hypothesis is essential, since a non-unital algebra may have too few unitaries for its unit ball to be generated in this way. The result also says nothing about preserving convex decompositions under a given map; it only supplies enough unit-ball approximants for norm estimates. [example: Unit Ball Control For A Unital Positive Map] Let $\Phi:A\to B$ be a unital positive map between unital $C^*$-algebras, and fix a unitary $u\in A$. The restriction of $\Phi$ to the abelian algebra $C^*(u)$ is a unital positive map out of a commutative $C^*$-algebra, hence is completely positive on $C^*(u)$. In $M_2(C^*(u))$, the element with entries $1,u,u^*,1$ is positive: in any faithful representation and for vectors $\xi,\eta$, \begin{align*} (\xi+u\eta,\xi+u\eta)=\|\xi\|^2+(u\eta,\xi)+(\xi,u\eta)+\|\eta\|^2\ge 0. \end{align*} Complete positivity therefore implies that the element of $M_2(B)$ with entries $1,\Phi(u),\Phi(u)^*,1$ is positive. We now extract the norm bound from this positive two-by-two element. Represent $B$ faithfully on a Hilbert space. For vectors $\xi,\eta$ and $t>0$, applying positivity to the pair $(t\xi,-\eta)$ gives \begin{align*} 0\le t^2\|\xi\|^2-2t\operatorname{Re}(\Phi(u)\eta,\xi)+\|\eta\|^2. \end{align*} Replacing $\xi$ by a scalar multiple of modulus $1$ makes $\operatorname{Re}(\Phi(u)\eta,\xi)=|(\Phi(u)\eta,\xi)|$. Since the displayed quadratic in $t$ is nonnegative for every $t>0$, its discriminant is nonpositive, so \begin{align*} |(\Phi(u)\eta,\xi)|^2\le \|\xi\|^2\|\eta\|^2. \end{align*} Taking the supremum over unit vectors $\xi$ gives $\|\Phi(u)\eta\|\le \|\eta\|$, and then taking the supremum over unit vectors $\eta$ gives $\|\Phi(u)\|\le 1$. Before passing to limits, note that $\Phi$ is norm-continuous. If $h=h^*$ and $\|h\|\le r$, then $-r1\le h\le r1$, so positivity and unitality give $-r1\le \Phi(h)\le r1$, hence $\|\Phi(h)\|\le r$. For $x=h+ik$ with $h=h^*$ and $k=k^*$, this gives \begin{align*} \|\Phi(x)\|\le \|\Phi(h)\|+\|\Phi(k)\|\le \|h\|+\|k\|\le 2\|x\|. \end{align*} Now let $a\in A$ satisfy $\|a\|\le 1$. By the *Russo-Dye theorem*, there is a net of convex combinations of unitaries \begin{align*} c_\alpha=\sum_{k=1}^{m_\alpha}t_{\alpha,k}u_{\alpha,k} \end{align*} with $t_{\alpha,k}\ge 0$, $\sum_{k=1}^{m_\alpha}t_{\alpha,k}=1$, and $c_\alpha\to a$ in norm. For each $\alpha$, linearity and the triangle inequality give \begin{align*} \|\Phi(c_\alpha)\|\le \sum_{k=1}^{m_\alpha}t_{\alpha,k}\|\Phi(u_{\alpha,k})\|\le \sum_{k=1}^{m_\alpha}t_{\alpha,k}=1. \end{align*} Since $\Phi$ is norm-continuous, $\Phi(c_\alpha)\to\Phi(a)$ in norm, so $\|\Phi(a)\|\le 1$. Therefore $\|\Phi\|\le 1$, while unitality gives $\|\Phi\|\ge \|\Phi(1)\|=\|1\|=1$. Hence $\|\Phi\|=1$: a unital positive map is controlled on the whole unit ball once it is controlled on unitaries. [/example] The chapter's main message is that normality is the compatibility condition between algebra and order. It is visible through suprema of projections, through monotone convergence of positive functionals, and through ultraweak continuity for maps. The predual packages all three viewpoints into a single Banach space attached canonically to every von Neumann algebra. The predual packages normality, positivity, and ultraweak continuity into a single framework, and that framework is what makes projection comparison effective. With normal maps and functionals under control, the course can now compare projections via partial isometries and central decompositions in the Murray-von Neumann sense. # 5. Murray-von Neumann Equivalence and Comparison This chapter turns projections from passive invariant subspaces into objects that can be compared. It assumes the earlier material on von Neumann algebras, projections, partial isometries, strong operator limits, and corners; it uses central projections in the comparison theorem, with the center developed systematically in Chapter 6. In a von Neumann algebra, the right notion of "same size" is not equality of Hilbert space dimension but equivalence through a partial isometry belonging to the algebra. The main questions are: when does one projection fit inside another, when does a projection contain a copy of itself as a proper subprojection, and how much comparison remains after passing to the centre? ## Comparing Projections by Partial Isometries How should two closed subspaces be compared when only the operators in a fixed von Neumann algebra are allowed? In $B(H)$, Hilbert space dimension decides unitary equivalence of projections. Inside a smaller von Neumann algebra $M \subseteq B(H)$, the implementing operator must lie in $M$, so the comparison records both the subspace size and the internal symmetries of $M$. [definition: Murray-von Neumann Equivalence] Let $M \subseteq B(H)$ be a von Neumann algebra, and let $p,q \in M$ be projections. The projections $p$ and $q$ are Murray-von Neumann equivalent in $M$, written $p \sim q$, if there exists a partial isometry $v \in M$ such that \begin{align*} v^*v &= p, & vv^* &= q. \end{align*} [/definition] The equations say that $v$ has initial projection $p$ and final projection $q$. Thus $v$ implements an isometric identification from $pH$ onto $qH$, but the point is that the implementing partial isometry is required to belong to $M$. [example: Rank Comparison in $B(H)$] Let $H$ be a Hilbert space and let $p,q \in B(H)$ be projections. We show that $p \sim q$ in $B(H)$ exactly when $\dim pH=\dim qH$. Assume first that $\dim pH=\dim qH$. Choose a Hilbert space unitary $u_0:pH\to qH$, and define $v:H\to H$ by \begin{align*} v\xi=u_0(p\xi). \end{align*} Since $p$ is a bounded projection and $u_0$ is an isometry on $pH$, the operator $v$ is bounded, hence $v\in B(H)$. If $\eta\in pH$, then $v\eta=u_0\eta$, while if $\eta\in (pH)^\perp$, then $p\eta=0$ and hence $v\eta=0$. Thus $v$ is the extension of $u_0$ by zero on $(pH)^\perp$. For $\xi,\eta\in H$, \begin{align*} \langle v\xi,v\eta\rangle=\langle u_0p\xi,u_0p\eta\rangle=\langle p\xi,p\eta\rangle=\langle \xi,p\eta\rangle, \end{align*} so $v^*v=p$. Also $vH=u_0(pH)=qH$, and $v$ is zero on $(pH)^\perp$, so the final projection of $v$ is the orthogonal projection onto $qH$, namely $q$. Therefore $vv^*=q$, and $p\sim q$ in $B(H)$. Conversely, suppose $p\sim q$ in $B(H)$. Then there is a partial isometry $v\in B(H)$ with $v^*v=p$ and $vv^*=q$. If $\xi\in pH$, then $p\xi=\xi$, so \begin{align*} \|v\xi\|^2=\langle v\xi,v\xi\rangle=\langle v^*v\xi,\xi\rangle=\langle p\xi,\xi\rangle=\|\xi\|^2. \end{align*} Thus $v$ is isometric on $pH$. Its range on $pH$ is $v(pH)=vH$, because $v\xi=v(p\xi)$ for every $\xi\in H$. Since $vv^*=q$, the range projection of $v$ is $q$, so $v(pH)=qH$. Hence $v|_{pH}:pH\to qH$ is a Hilbert space unitary, and therefore $\dim pH=\dim qH$. Thus in $B(H)$, Murray-von Neumann equivalence of projections is exactly equality of Hilbert space dimension. [/example] Equivalence gives a symmetric notion of equal size. To express that one projection fits inside another, we compare a projection with a subprojection of the other one. [definition: Murray-von Neumann Subequivalence] Let $M \subseteq B(H)$ be a von Neumann algebra, and let $p,q \in M$ be projections. The projection $p$ is Murray-von Neumann subequivalent to $q$ in $M$, written $p \precsim q$, if there exists a projection $q_0 \in M$ with $q_0 \le q$ and $p \sim q_0$. [/definition] Subequivalence is the projection analogue of an order relation, but it is defined using the internal geometry of $M$. Before it can be used for classification, we need to know that equivalence and subequivalence have the formal behaviour expected from equality and order. [quotetheorem:9278] [citeproof:9278] The theorem justifies writing comparison statements without naming the implementing partial isometries every time, but its hypotheses are part of the content. The partial isometries must lie in the same von Neumann algebra $M$; two projections may have Hilbert spaces of the same dimension and still fail to be equivalent if $M$ does not contain an operator carrying one range onto the other. The result also does not say that $p \precsim q$ and $q \precsim p$ are automatically the same as equality of projections; it only prepares the comparison relation whose antisymmetry must still be proved after quotienting by equivalence. The next result is the projection version of Schroeder-Bernstein: mutual embeddability forces equivalence. [quotetheorem:9279] [citeproof:9279] This theorem is the reason subequivalence behaves like an order after passing to equivalence classes. Both assumptions are essential: a single embedding $p\precsim q$ only says that $p$ fits into $q$, and gives no way to recover the part of $q$ outside the matched subprojection. The ambient von Neumann algebra is also essential, because the theorem uses strong operator limits of partial isometries inside $M$; mutual Hilbert-space embeddability of $pH$ and $qH$ would only compare external dimensions and would not produce an implementing partial isometry in a prescribed smaller algebra. The theorem therefore identifies the exact point at which internal comparison becomes an order, and it prepares the language for finiteness, where the obstruction is the possibility that a projection contains a proper copy of itself. ## Finite and Infinite Projections Can a projection be the same size as a proper subprojection of itself? In finite-dimensional linear algebra this cannot happen, but infinite-dimensional Hilbert spaces admit shifts that identify the whole space with a proper closed subspace. Von Neumann algebras measure this phenomenon internally, using partial isometries in the algebra. [definition: Finite Projection] Let $M \subseteq B(H)$ be a von Neumann algebra, and let $p \in M$ be a projection. The projection $p$ is finite in $M$ if, whenever $q \in M$ is a projection with $q \le p$ and $q \sim p$, then $q=p$. [/definition] A finite projection is therefore one that cannot be compressed onto a proper subprojection by an allowed partial isometry. This motivates naming the opposite phenomenon: a projection is infinite when such a proper self-copy exists inside the algebra. [definition: Infinite Projection] Let $M \subseteq B(H)$ be a von Neumann algebra, and let $p \in M$ be a projection. The projection $p$ is infinite in $M$ if there exists a projection $q \in M$ such that $q < p$ and $q \sim p$. [/definition] The two definitions are complementary, but there is a stronger condition that matters in type classification. Proper infiniteness asks for two disjoint copies of the projection inside itself. [definition: Properly Infinite Projection] Let $M \subseteq B(H)$ be a von Neumann algebra, and let $p \in M$ be a projection. The projection $p$ is properly infinite in $M$ if there exist orthogonal projections $p_1,p_2 \in M$ such that \begin{align*} p_1 &\le p, & p_2 &\le p, & p_1 \sim p, & p_2 \sim p. \end{align*} [/definition] This convention includes the zero projection, since $p_1=p_2=0$ witnesses proper infiniteness for $p=0$. Some authors reserve the phrase for nonzero projections; when nonzero behaviour matters, the notes will say so explicitly. Proper infiniteness is stronger than infiniteness for a nonzero projection, though the distinction can collapse in many factor situations. It is the operator-algebraic version of the Hilbert space fact that a separably infinite-dimensional space contains two orthogonal copies of itself. [example: The Unilateral Shift in $B(\ell^2)$] Let $H=\ell^2(\mathbb N)$ with standard orthonormal basis $(e_n)_{n\ge 1}$, and define $S\in B(H)$ by $Se_n=e_{n+1}$. For a finitely supported vector $\xi=\sum_{n\ge 1} a_n e_n$, we have \begin{align*} \|S\xi\|^2=\left\|\sum_{n\ge 1}a_n e_{n+1}\right\|^2=\sum_{n\ge 1}|a_n|^2=\|\xi\|^2. \end{align*} By density of finitely supported vectors, $S$ extends to an isometry on $H$, so $S^*S=I$. The range of $S$ is the closed span of $\{e_2,e_3,\dots\}$, whose orthogonal projection is $I-e_1\otimes e_1$, where $(e_1\otimes e_1)\xi=\langle \xi,e_1\rangle e_1$. Hence \begin{align*} SS^*=I-e_1\otimes e_1. \end{align*} Thus $S$ is a partial isometry in $B(H)$ with initial projection $I$ and final projection $I-e_1\otimes e_1$, so $I\sim I-e_1\otimes e_1$ in $B(H)$. Since $(I-e_1\otimes e_1)e_1=0$ while $Ie_1=e_1$, the projection $I-e_1\otimes e_1$ is a proper subprojection of $I$. Therefore $I$ is infinite in $B(H)$. To see proper infiniteness, define isometries $V_1,V_2\in B(H)$ on the basis by \begin{align*} V_1e_n=e_{2n-1}. \end{align*} \begin{align*} V_2e_n=e_{2n}. \end{align*} The same norm computation gives $V_1^*V_1=I$ and $V_2^*V_2=I$. Their range projections are the projections $p_1$ onto $\overline{\operatorname{span}}\{e_1,e_3,e_5,\dots\}$ and $p_2$ onto $\overline{\operatorname{span}}\{e_2,e_4,e_6,\dots\}$. These subspaces are orthogonal, so $p_1p_2=0$, and both projections satisfy $p_1\le I$ and $p_2\le I$. Since \begin{align*} V_1^*V_1=I,\quad V_1V_1^*=p_1 \end{align*} and \begin{align*} V_2^*V_2=I,\quad V_2V_2^*=p_2, \end{align*} we have $p_1\sim I$ and $p_2\sim I$. Thus $I$ contains two orthogonal subprojections each Murray-von Neumann equivalent to $I$, so $I$ is properly infinite in $B(\ell^2(\mathbb N))$. [/example] The shift example shows how infiniteness is detected by a proper self-copy. To use finite projections in corners and comparison arguments, we need intrinsic tests that do not depend on the particular way a self-copy might be presented. [quotetheorem:9280] [citeproof:9280] This characterization makes finite projections stable under working inside corners, but the restriction $e,f\le p$ is essential: finiteness of $p$ only controls comparison inside the corner supported by $p$, not arbitrary projections elsewhere in $M$. For instance, in $B(\ell^2(\mathbb N))$ a rank-one projection $p$ is finite, while the identity projection is infinite; comparison phenomena outside $pB(H)p$ are invisible to the finiteness of $p$. The theorem is therefore not a classification of all projections in $M$, but a local test for whether the piece of the algebra cut out by $p$ behaves like finite-dimensional rank theory. This is why matrix algebras behave like finite-dimensional rank theory while $B(H)$ on an infinite-dimensional Hilbert space does not. [example: Matrix Algebras Over L Infinity] Let $(X,\mathcal F,\mu)$ be a measure space and let $M=M_n(L^\infty(X,\mu))$ act on $L^2(X,\mu;\mathbb C^n)$ by pointwise matrix multiplication. A projection $p\in M$ is an essentially bounded [measurable function](/page/Measurable%20Function) $x\mapsto p(x)\in M_n(\mathbb C)$ satisfying $p(x)^2=p(x)=p(x)^*$ for a.e. $x$, so each $p(x)$ is the orthogonal projection onto a subspace of $\mathbb C^n$. First suppose $p\sim q$ in $M$, implemented by $v\in M$ with $v^*v=p$ and $vv^*=q$. Evaluating pointwise gives, for a.e. $x$, \begin{align*} v(x)^*v(x)=p(x),\qquad v(x)v(x)^*=q(x). \end{align*} For a finite matrix $A$, the matrices $A^*A$ and $AA^*$ have the same rank, because $\ker(A^*A)=\ker(A)$ and $\operatorname{rank}(A^*A)=n-\dim\ker(A) = \operatorname{rank}(A) = \operatorname{rank}(AA^*)$. Applying this to $A=v(x)$ gives \begin{align*} \operatorname{rank}p(x)=\operatorname{rank}(v(x)^*v(x))=\operatorname{rank}(v(x)v(x)^*)=\operatorname{rank}q(x) \end{align*} for a.e. $x$. Conversely, assume $\operatorname{rank}p(x)=\operatorname{rank}q(x)$ for a.e. $x$. On the measurable set where this common rank is $k$, choose measurable orthonormal bases $u_1(x),\dots,u_k(x)$ for $p(x)\mathbb C^n$ and $w_1(x),\dots,w_k(x)$ for $q(x)\mathbb C^n$; in finite dimension these are obtained by applying Gram-Schmidt to the coordinate columns of $p(x)$ and $q(x)$ on the measurable rank strata. Define \begin{align*} v(x)\eta=\sum_{j=1}^k \langle \eta,u_j(x)\rangle w_j(x). \end{align*} Then \begin{align*} v(x)^*\eta=\sum_{j=1}^k \langle \eta,w_j(x)\rangle u_j(x). \end{align*} Thus, for $\eta\in\mathbb C^n$, \begin{align*} v(x)^*v(x)\eta=\sum_{j=1}^k \langle \eta,u_j(x)\rangle u_j(x)=p(x)\eta. \end{align*} Similarly, \begin{align*} v(x)v(x)^*\eta=\sum_{j=1}^k \langle \eta,w_j(x)\rangle w_j(x)=q(x)\eta. \end{align*} The entries of $v(x)$ are measurable and bounded by $1$, so $v\in M$, and therefore $p\sim q$. The same pointwise argument gives subequivalence. If $p\precsim q$, then $p\sim q_0$ for some projection $q_0\le q$, so \begin{align*} \operatorname{rank}p(x)=\operatorname{rank}q_0(x)\le \operatorname{rank}q(x) \end{align*} for a.e. $x$. Conversely, if $\operatorname{rank}p(x)\le \operatorname{rank}q(x)$ a.e., choose the first $\operatorname{rank}p(x)$ vectors in a measurable orthonormal basis of $q(x)\mathbb C^n$ and let $q_0(x)$ be the projection onto their span. Then $q_0\le q$, $\operatorname{rank}q_0(x)=\operatorname{rank}p(x)$ a.e., and the construction above gives $p\sim q_0$. Hence $p\precsim q$ exactly when $\operatorname{rank}p(x)\le \operatorname{rank}q(x)$ a.e. Projection comparison in $M_n(L^\infty(X,\mu))$ is therefore encoded by measurable rank functions $X\to\{0,1,\dots,n\}$, modulo equality a.e. [/example] The example shows that comparison need not reduce to a single number outside factors. The centre carries pointwise data, and the next section isolates how central projections control this behaviour. ## Central Support and Comparison in Factors What part of the algebra does a projection see? A projection may be nonzero but live entirely on a central summand, and comparison across disjoint central summands is impossible. For example, in a direct sum $M=M_1\oplus M_2$, the projections $p=(1,0)$ and $q=(0,1)$ are both nonzero, but there is no partial isometry $v\in M$ with $v^*v=p$ and $vv^*=q$, since multiplication is coordinatewise and the two projections occupy different central summands. Central support records the smallest central projection that contains the projection. [definition: Central Support] Let $M \subseteq B(H)$ be a von Neumann algebra with center $Z(M)$, and let $p \in M$ be a projection. The central support of $p$, denoted $z(p)$, is the smallest central projection $z \in Z(M)$ such that $p \le z$. [/definition] The existence of $z(p)$ follows because arbitrary meets of central projections exist in a von Neumann algebra. Central support measures the central summand generated by $p$, not the ordinary range size of $pH$. [remark: Central Support in a Factor] If $M$ is a factor, then $Z(M)=\mathbb C I$, so the only central projections are $0$ and $I$. Hence every nonzero projection in a factor has central support $I$. [/remark] The remark says that in a factor there is no central location where two nonzero projections can avoid each other. This removes the main obstruction to comparison, and it leads to a total comparison theorem for projections. [quotetheorem:9281] [citeproof:9281] The factor hypothesis is essential: if $M=M_1\oplus M_2$, then $p=(1,0)$ and $q=(0,1)$ have disjoint central supports, so neither projection can contain an internal copy of the other. Thus total comparison is not a statement about all von Neumann algebras, but about algebras with no nontrivial central decomposition. Even in a factor, the theorem only gives one comparison direction; it does not decide whether the smaller projection is equivalent to the larger one, finite, or properly infinite. The result is the structural reason that factors admit type classifications by projection behaviour, while in a general von Neumann algebra the statement must be refined by central projections, splitting the algebra into central regions where one comparison direction holds. [quotetheorem:9282] [citeproof:9282] Central comparison contains factor comparison as the case where the centre has only $0$ and $I$, and this shows why the central projection in the statement is necessary rather than decorative. Without splitting by $z$, projections on different central summands can be incomparable, so no single global direction of comparison should be expected. The theorem also has a limitation: it gives a central partition on which one projection embeds into the other, but it does not assign numerical sizes or produce a trace-like dimension function. Its forward use is to separate the algebra into regions where comparison behaves factorwise, which is exactly the setting needed for direct-integral intuition, commutative centres, and later trace and weight constructions. [example: Central Comparison by Rank Functions] In $M_n(L^\infty(X,\mu))$, central projections are scalar projection-valued functions, so the set \begin{align*} E=\{x\in X:r_p(x)\le r_q(x)\} \end{align*} defines the central projection $z=\chi_E I_n$. For a.e. $x$, \begin{align*} (zp)(x)=\chi_E(x)p(x) \end{align*} and \begin{align*} (zq)(x)=\chi_E(x)q(x). \end{align*} Thus, if $x\in E$, then $\operatorname{rank}(zp)(x)=r_p(x)\le r_q(x)=\operatorname{rank}(zq)(x)$, while if $x\notin E$, both ranks are $0$. Hence \begin{align*} \operatorname{rank}(zp)(x)\le \operatorname{rank}(zq)(x) \end{align*} for a.e. $x$, and the rank comparison criterion for projections in $M_n(L^\infty(X,\mu))$ gives $zp\precsim zq$. For the complementary central projection $I-z=\chi_{X\setminus E}I_n$, we similarly have \begin{align*} ((I-z)q)(x)=\chi_{X\setminus E}(x)q(x) \end{align*} and \begin{align*} ((I-z)p)(x)=\chi_{X\setminus E}(x)p(x). \end{align*} If $x\in X\setminus E$, then $r_p(x)>r_q(x)$, so $r_q(x)\le r_p(x)$; if $x\in E$, both complementary ranks are $0$. Therefore \begin{align*} \operatorname{rank}((I-z)q)(x)\le \operatorname{rank}((I-z)p)(x) \end{align*} for a.e. $x$, and the same rank comparison criterion gives $(I-z)q\precsim (I-z)p$. Thus the abstract central comparison projection is, in this algebra, exactly the characteristic function of the region where the rank of $p$ is at most the rank of $q$. [/example] The chapter's main lesson is that projections in a von Neumann algebra form an internal dimension theory. Partial isometries define equivalence, subequivalence defines comparison, finite projections forbid self-compression, and central support records where comparison is allowed to occur. In the next part of the course, traces and weights turn this comparison theory into numerical and extended-valued dimension functions. The same pattern appears outside operator algebras whenever equality is too rigid and embedding is the useful comparison. Cardinal arithmetic compares sets by injections before passing to bijection, algebraic geometry decomposes objects over central or idempotent pieces of a base ring, and measure theory separates spaces into regions before assigning numerical size. Murray-von Neumann comparison is the noncommutative version of this idea: partial isometries replace bijections, central projections replace measurable regions, and traces later play the role of dimension or measure. Projection comparison clarifies how size can be measured without commutativity, but the result also shows that the center governs the algebra's global splitting. The next chapter isolates that central structure and studies factors, where the decomposition disappears and the remaining theory becomes genuinely indecomposable. # 6. Factors and Centers This chapter asks how a von Neumann algebra splits when some observables commute with everything. In earlier chapters, commutants measured symmetry and closure; now the center measures internal decomposition. The guiding principle is that central projections label independent sectors, while a factor is an algebra with no non-scalar central observables. ## The Center and the Factor Condition What information remains visible to every part of a von Neumann algebra? If an operator commutes with every observable in $M$, then it cannot distinguish noncommuting internal directions, but it may still distinguish direct summands. This leads to the center. [definition: Center of a Von Neumann Algebra] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra. The center of $M$ is \begin{align*} Z(M) := M \cap M' = \{x \in M : xy = yx \text{ for all } y \in M\}. \end{align*} [/definition] The center is itself an abelian von Neumann algebra. It is the part of $M$ that behaves like a classical algebra of labels rather than a genuinely noncommutative algebra of observables. [example: Matrix Algebra Has Scalar Center] Let $M=M_n(\mathbb C)$ act on $\mathbb C^n$, and write $a=(a_{kl})_{k,l=1}^n$. For the matrix unit $e_{ij}$, the $(k,l)$-entry of $ae_{ij}$ is \begin{align*} (ae_{ij})_{kl}=\sum_{m=1}^n a_{km}(e_{ij})_{ml}=a_{ki}\delta_{jl}. \end{align*} The $(k,l)$-entry of $e_{ij}a$ is \begin{align*} (e_{ij}a)_{kl}=\sum_{m=1}^n (e_{ij})_{km}a_{ml}=\delta_{ki}a_{jl}. \end{align*} If $a$ commutes with every $e_{ij}$, then for all $i,j,k,l$ we have \begin{align*} a_{ki}\delta_{jl}=\delta_{ki}a_{jl}. \end{align*} Taking $l=j$ and $k\ne i$ gives $a_{ki}=0$, so every off-diagonal entry of $a$ is zero. Taking $k=i$ and $l=j$ gives $a_{ii}=a_{jj}$, so all diagonal entries are equal. Thus $a=\lambda I_n$ for some $\lambda\in\mathbb C$. Conversely, every scalar matrix $\lambda I_n$ commutes with every element of $M_n(\mathbb C)$, so \begin{align*} Z(M_n(\mathbb C))=\mathbb C I_n. \end{align*} The finite matrix algebra therefore has no non-scalar central observables. [/example] The matrix example shows that a center may collapse to scalars, but in larger algebras the center can contain genuine projections. To turn the center into a decomposition tool, we need to single out its projection-valued elements, since projections cut Hilbert spaces into reducing subspaces. [definition: Central Projection] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra. A projection $z \in M$ is central if $z \in Z(M)$. [/definition] A central projection cuts the Hilbert space as $H=zH\oplus(1-z)H$, but a Hilbert-space splitting alone does not yet say that the algebra itself has split into independent pieces. The obstruction is invariance under multiplication from both sides: if the projection were not central, operators in $M$ could mix the two subspaces and the would-be summands would not be two-sided von Neumann algebra pieces. Centrality removes this obstruction and should turn the geometric splitting into an algebraic direct-sum decomposition of $M$. [quotetheorem:9283] [citeproof:9283] This theorem explains why the center is a decomposition device: a nonzero proper central projection is exactly the data needed to split $M$ into two independent von Neumann algebras. The centrality hypothesis is essential: if $p$ is a noncentral projection in $M_2(\mathbb C)$, then $pM$ is not a two-sided von Neumann algebra summand, because multiplication by matrices that mix $p\mathbb C^2$ with $(1-p)\mathbb C^2$ leaves the corner. The theorem also does not say that every Hilbert-space decomposition gives an algebra decomposition; it only applies to reducing decompositions detected inside $Z(M)$. This motivates the following definition, which names the case where all central splitting has been removed. [definition: Factor] A von Neumann algebra $M \subseteq \mathcal{L}(H)$ is a factor if \begin{align*} Z(M)=\mathbb C I. \end{align*} [/definition] The definition is stated using all central operators, but the preceding decomposition theorem suggests a projection-level test. The obstruction is that a non-scalar central self-adjoint operator need not itself be a projection, so it is not immediately visible as a direct summand. Spectral theory resolves this obstruction by turning central self-adjoint operators into central spectral projections. The next criterion is useful because projection lattices are the main language of von Neumann algebra structure. [quotetheorem:9284] [citeproof:9284] The spectral-theoretic argument works inside the abelian von Neumann algebra $Z(M)$. The von Neumann algebra hypothesis is needed here: a general $*$-subalgebra need not contain the spectral projections of its self-adjoint elements, so projection data may fail to recover central operator data. For example, the polynomial algebra generated by a non-scalar self-adjoint matrix can contain central self-adjoint elements without being closed under all spectral projections until one passes to its finite-dimensional closure. The result also does not classify factors; it only gives a test for when the center is trivial. The next examples show the two extremes: matrix and full operator algebras satisfy the test, while abelian algebras usually fail it completely. [example: Abelian Von Neumann Algebras Are the Opposite Extreme] Let $M_f$ denote multiplication by $f\in L^\infty(X,\mu)$ on $L^2(X,\mu)$. For $f,g\in L^\infty(X,\mu)$ and $\xi\in L^2(X,\mu)$, multiplication is pointwise, so \begin{align*} (M_fM_g\xi)(t)=f(t)g(t)\xi(t)=g(t)f(t)\xi(t)=(M_gM_f\xi)(t) \end{align*} for almost every $t$. Hence every two elements of $M$ commute, so $M\subseteq M'$. In the standard multiplication representation, the multiplication algebra is maximal abelian: every bounded operator on $L^2(X,\mu)$ commuting with all multiplication operators is itself a multiplication operator. Thus $M'=M$, and by the definition of the center, \begin{align*} Z(M)=M\cap M'=M\cap M=M. \end{align*} Therefore $M$ is a factor exactly when $M=\mathbb C I$, equivalently when every essentially bounded measurable function is constant modulo null sets. If $E\subseteq X$ has both $E$ and $X\setminus E$ non-null, then $M_{\mathbf 1_E}$ is a projection in $Z(M)$ because $\mathbf 1_E^2=\mathbf 1_E$, but it is neither $0$ nor $I$. Conversely, if the measure algebra has no nontrivial measurable subsets, then the level sets of any $f\in L^\infty(X,\mu)$ force $f$ to be essentially constant. Thus abelian von Neumann algebras are factors only in the one-dimensional measure-algebra case. [/example] ## Direct Sums and Superselection Sectors How should we interpret a nontrivial center? In the operator-algebraic formalism, central projections describe mutually exclusive sectors that cannot be mixed by observables in $M$. This is the algebraic version of decomposing a representation into superselection sectors. [definition: Direct Sum of Von Neumann Algebras] Let $(M_i)_{i\in I}$ be von Neumann algebras acting on Hilbert spaces $(H_i)_{i\in I}$. Their bounded direct sum is \begin{align*} \bigoplus_{i\in I} M_i := \left\{(x_i)_{i\in I}: x_i\in M_i,\ \sup_{i\in I}\|x_i\| < \infty\right\}, \end{align*} acting diagonally on $\bigoplus_{i\in I}H_i$. [/definition] For a finite direct sum, central projections are especially concrete: they are sums of identity projections on selected summands. This makes the sector interpretation visible without measure-theoretic central decomposition. [example: Two Superselection Sectors] Let $M=M_2(\mathbb C)\oplus M_3(\mathbb C)$ act diagonally on $\mathbb C^2\oplus \mathbb C^3$, so \begin{align*} (a,b)(\xi,\eta)=(a\xi,b\eta) \end{align*} for $a\in M_2(\mathbb C)$, $b\in M_3(\mathbb C)$, $\xi\in\mathbb C^2$, and $\eta\in\mathbb C^3$. Put $z=(I_2,0)$. Then \begin{align*} z^2=(I_2,0)(I_2,0)=(I_2^2,0^2)=(I_2,0)=z. \end{align*} Also \begin{align*} z^*=(I_2^*,0^*)=(I_2,0)=z. \end{align*} Thus $z$ is a projection. For any observable $(a,b)\in M$, \begin{align*} z(a,b)=(I_2a,0b)=(a,0). \end{align*} On the other hand, \begin{align*} (a,b)z=(aI_2,b0)=(a,0). \end{align*} Hence $z(a,b)=(a,b)z$ for every $(a,b)\in M$, so $z\in Z(M)$. The range of $z$ is the first summand, since \begin{align*} z(\xi,\eta)=(I_2\xi,0\eta)=(\xi,0). \end{align*} The complementary projection is $1-z=(0,I_3)$, and \begin{align*} (1-z)(\xi,\eta)=(0\xi,I_3\eta)=(0,\eta). \end{align*} Finally, an observable $(a,b)$ preserves each summand because \begin{align*} (a,b)(\xi,0)=(a\xi,0) \end{align*} and \begin{align*} (a,b)(0,\eta)=(0,b\eta). \end{align*} So no element of $M$ sends a vector from $\mathbb C^2$ into $\mathbb C^3$ or from $\mathbb C^3$ into $\mathbb C^2$; the central projection $z$ is exactly the label for the first superselection sector. [/example] The example shows that central projections in a direct sum are sector indicators. To compute all such labels, and to see what happens when each summand is already a factor, we need the center of a direct sum. [quotetheorem:9285] [citeproof:9285] The finite or discrete direct sum picture is the model for the more general central decomposition theory. The boundedness condition in the direct sum is essential: without $\sup_i\|x_i\|<\infty$, the coordinatewise family need not define a bounded operator on $\bigoplus_i H_i$. The theorem also does not say that every von Neumann algebra is literally a discrete direct sum of factors; a diffuse center such as $L^\infty([0,1])$ leads instead to a direct-integral decomposition indexed by a measure space. This limitation is exactly why central projections are only the first layer of the general central decomposition theory. [explanation: Superselection Language] A central projection $z$ represents a yes-no question whose answer is compatible with every observable in $M$. If the answer is yes, the physical state lives in the sector $zH$; if no, it lives in $(1-z)H$. Since all observables commute with $z$, no observable in $M$ transfers vectors between these two sectors. The center is therefore the classical algebra of sector labels, while each factor summand contains the noncommutative observables internal to a sector. [/explanation] The superselection picture raises a local question about an ordinary projection $p\in M$: which central sector does $p$ occupy? The obstruction is that $p$ need not itself be central, so it may describe a subspace inside a sector rather than a whole sector. In $M_2(\mathbb C)\oplus M_3(\mathbb C)$, a rank-one projection in the first block is not a central projection, but it should still be recognised as living entirely in the first summand. To answer this, we attach to $p$ the smallest central projection that contains it. [definition: Central Carrier] Let $M$ be a von Neumann algebra and let $p\in M$ be a projection. The central carrier, also called the central support in Chapter 5, of $p$ is the projection \begin{align*} c(p):=\bigwedge\{z\in Z(M): z \text{ is a projection and } p\le z\}. \end{align*} [/definition] The meet exists because projections in a von Neumann algebra form a complete lattice, and central projections form a complete sublattice. The definition, however, is order-theoretic and can be hard to recognize in computations: it describes $c(p)$ as an infimum over all central projections above $p$. A more usable test asks which central sectors are disjoint from $p$. If a central projection annihilates $p$, it should also annihilate the entire central sector generated by $p$, and conversely. [quotetheorem:9286] [citeproof:9286] This theorem says that $c(p)$ is detected by the central projections that do not annihilate $p$. The centrality of $z$ is essential in the annihilator test: a noncentral projection can be orthogonal to $p$ without annihilating the whole central sector generated by $p$, as happens for two orthogonal rank-one projections inside $M_2(\mathbb C)$. The theorem also does not say that $p$ is equivalent to $c(p)$, or that $p$ fills the whole sector as a projection; it only records the smallest central support in which $p$ lives. In a factor, every nonzero projection has central carrier $1$, so every nonzero projection meets the unique central sector, a fact used later when comparing projections by Murray-von Neumann equivalence rather than by central support alone. [example: Central Carrier in a Direct Sum] Let $M=M_2(\mathbb C)\oplus M_3(\mathbb C)$ and let $p=(e_{11},0)$. Since $e_{11}^2=e_{11}$ and $e_{11}^*=e_{11}$, we have \begin{align*} p^2=(e_{11}^2,0^2)=(e_{11},0)=p \end{align*} and \begin{align*} p^*=(e_{11}^*,0^*)=(e_{11},0)=p. \end{align*} Thus $p$ is a projection. An element $(a,b)\in M$ is central exactly when $a$ commutes with every element of $M_2(\mathbb C)$ and $b$ commutes with every element of $M_3(\mathbb C)$. From the matrix-algebra center computation, this means \begin{align*} a=\lambda I_2 \text{ and } b=\mu I_3 \end{align*} for some $\lambda,\mu\in\mathbb C$. If $(a,b)$ is also a projection, then \begin{align*} (\lambda I_2,\mu I_3)^2=(\lambda^2 I_2,\mu^2 I_3)=(\lambda I_2,\mu I_3), \end{align*} so $\lambda^2=\lambda$ and $\mu^2=\mu$. Hence $\lambda,\mu\in\{0,1\}$, and the central projections are \begin{align*} (0,0),\ (I_2,0),\ (0,I_3),\ (I_2,I_3). \end{align*} For projections, $p\le z$ means $zp=p$. Checking the four central projections gives \begin{align*} (0,0)p=(0,0)\ne p, \end{align*} \begin{align*} (I_2,0)p=(I_2e_{11},0)=(e_{11},0)=p, \end{align*} \begin{align*} (0,I_3)p=(0e_{11},I_3 0)=(0,0)\ne p, \end{align*} and \begin{align*} (I_2,I_3)p=(I_2e_{11},I_3 0)=(e_{11},0)=p. \end{align*} Thus the central projections dominating $p$ are $(I_2,0)$ and $(I_2,I_3)$. Since \begin{align*} (I_2,0)(I_2,I_3)=(I_2,0), \end{align*} we have $(I_2,0)\le (I_2,I_3)$, so the smallest central projection dominating $p$ is \begin{align*} c(p)=(I_2,0). \end{align*} Thus a rank-one projection in the first matrix block has full central support in the first sector and no central support in the second. [/example] ## Prime Sources of Factors Which von Neumann algebras actually have trivial center? The first examples come from full operator algebras, then from group representations, and finally from tensor products. These examples supply the main families used later in type classification. [quotetheorem:9287] [citeproof:9287] This includes both $M_n(\mathbb C)$ and $\mathcal{L}(H)$ for infinite-dimensional $H$. The hypothesis $H\ne\{0\}$ only avoids the degenerate zero algebra, whose identity convention can obscure the statement of the center. The theorem is specific to the full operator algebra: the compact operators $\mathcal K(H)$ on an infinite-dimensional Hilbert space also have scalar multiplier center, but they are not a von Neumann algebra because they are not weakly closed. Thus the result identifies a basic source of factors, while later type theory separates finite matrix factors from the infinite full algebra by traces and projection comparison. [example: Infinite-Dimensional Full Algebra] Let $H=\ell^2(\mathbb N)$ with orthonormal basis $(e_m)_{m\ge 1}$. For each $n$, the operator \begin{align*} P_n\xi=\sum_{m=1}^n(\xi,e_m)e_m \end{align*} is a finite-rank projection in $\mathcal L(H)$. The projection onto the closed span of the even basis vectors, \begin{align*} P_{\mathrm{even}}\xi=\sum_{m=1}^\infty(\xi,e_{2m})e_{2m}, \end{align*} is an infinite-rank projection in $\mathcal L(H)$. The unilateral shift $S e_m=e_{m+1}$ also belongs to $\mathcal L(H)$, and it is not compact: for $i\ne j$, \begin{align*} \|S e_i-S e_j\|^2=\|e_{i+1}-e_{j+1}\|^2=1+1=2, \end{align*} so the bounded sequence $(S e_m)$ has no norm-convergent subsequence. Rank-one operators force central operators in $\mathcal L(H)$ to be scalar. If $T\in\mathcal L(H)$ commutes with every rank-one operator, fix $i,j$ and let $R_{ij}$ be defined by $R_{ij}e_j=e_i$ and $R_{ij}e_k=0$ for $k\ne j$. From $TR_{ij}=R_{ij}T$, applying both sides to $e_j$ gives \begin{align*} T e_i=TR_{ij}e_j=R_{ij}T e_j=(T e_j,e_j)e_i. \end{align*} Thus every basis vector is an eigenvector of $T$. Taking $i$ fixed and comparing the same formula for two choices of $j$ shows the eigenvalue is independent of $j$, so $T=\lambda I$ for some $\lambda\in\mathbb C$. Therefore $\mathcal L(H)$ has scalar center. The compact operators behave differently as a von Neumann-algebra example. By definition, $\mathcal K(H)$ is the norm closure of the finite-rank operators, but it is not weakly closed: the projections $P_n$ are finite rank, and for every $\xi\in H$, \begin{align*} \|\xi-P_n\xi\|^2=\sum_{m=n+1}^\infty |(\xi,e_m)|^2\to 0, \end{align*} so $P_n\to I$ strongly and hence weakly. But $I\notin\mathcal K(H)$, since the sequence $(I e_m)=(e_m)$ has no norm-convergent subsequence. Thus $\mathcal L(H)$ is a factor, while $\mathcal K(H)$ is not a von Neumann algebra at all; scalar central behaviour is not enough unless the algebra is weakly closed. [/example] Full operator algebras give factors from Hilbert-space linear algebra. A different source comes from discrete groups, where the obstruction to being a factor is the presence of finite nonidentity conjugacy classes. [definition: ICC Group] A discrete group $G$ is ICC if every [conjugacy class](/page/Conjugacy%20Class) other than $\{e\}$ is infinite. [/definition] The ICC condition removes finite conjugacy-class elements, which would otherwise produce central vectors in the group von Neumann algebra. To state the operator algebra attached to $G$, we represent $G$ by left translations on $\ell^2(G)$ and close in the bicommutant sense. [definition: Group Von Neumann Algebra] Let $G$ be a discrete group and let $\lambda:G\to \mathcal{L}(\ell^2(G))$ be the left regular representation. The group von Neumann algebra of $G$ is \begin{align*} L(G):=\{\lambda_g:g\in G\}''. \end{align*} [/definition] The double commutant definition places $L(G)$ within the framework developed earlier. The obstruction to factoriality is a finite nonidentity conjugacy class: a conjugacy-invariant coefficient function supported on such a class can produce a non-scalar central element. The ICC condition removes exactly this obstruction by forcing every nonidentity conjugacy class to be too large to support a nonzero square-summable constant coefficient. This motivates the following criterion, because central elements in $L(G)$ must be invariant under conjugation by the group unitaries. [quotetheorem:9288] [citeproof:9288] This theorem is a bridge between algebraic group structure and operator-algebraic factor theory. The ICC hypothesis is essential: if $G$ is abelian and nontrivial, every conjugacy class is a singleton and $L(G)$ is abelian, so its center is all of $L(G)$ rather than just the scalars. More generally, finite nonidentity conjugacy classes can contribute non-scalar central elements through class-supported Fourier coefficients. The theorem does not say that all group factors arise from ICC groups in a classification sense; it only gives a powerful sufficient condition for the left regular construction. It produces many type $\mathrm{II}_1$ factors once traces and Murray-von Neumann equivalence are introduced. [example: Free Group Factor] Let $\mathbb F_2=\langle a,b\rangle$, and let $w\ne e$ be a reduced word of length $r\ge 1$. Choose a letter $c\in\{a,a^{-1},b,b^{-1}\}$ such that $c^{-1}$ is not the first letter of $w$ and $c$ is not the last letter of $w$; this is possible because at most two of the four letters are forbidden. Then no cancellation occurs in \begin{align*} c^nwc^{-n} \end{align*} for any $n\ge 1$: the left boundary does not cancel because the first letter of $w$ is not $c^{-1}$, and the right boundary does not cancel because the last letter of $w$ is not $c$. Hence $c^nwc^{-n}$ is a reduced word of length \begin{align*} n+r+n=r+2n. \end{align*} If $m\ne n$, then $r+2m\ne r+2n$, so the reduced words $c^mwc^{-m}$ and $c^nwc^{-n}$ are distinct. Thus every nonidentity element of $\mathbb F_2$ has infinitely many conjugates, so $\mathbb F_2$ is ICC. By the *[ICC Group Factor Criterion](/theorems/9288)*, $L(\mathbb F_2)$ is a factor. By contrast, write $\mathbb Z$ additively. For $m,n\in\mathbb Z$, conjugation gives \begin{align*} m+n-m=n. \end{align*} Thus every conjugacy class in $\mathbb Z$ is a singleton, so $\mathbb Z$ is not ICC unless the group is trivial. Also the left regular unitaries commute: \begin{align*} \lambda_m\lambda_n=\lambda_{m+n}=\lambda_{n+m}=\lambda_n\lambda_m. \end{align*} Therefore $L(\mathbb Z)$ is abelian, so its center is all of $L(\mathbb Z)$ rather than just the scalars. The ICC test separates the free-group example, which gives a genuinely noncommutative factor, from the commutative group von Neumann algebra of $\mathbb Z$. [/example] The full-operator and group examples give individual factors. To build new examples from old ones, we need an operation that combines two von Neumann algebras while respecting their concrete Hilbert-space representations. [definition: Spatial Tensor Product] Let $M\subseteq \mathcal{L}(H)$ and $N\subseteq \mathcal{L}(K)$ be von Neumann algebras. Their spatial [tensor product](/page/Tensor%20Product) is \begin{align*} M\overline{\otimes}N := (M\odot N)'' \subseteq \mathcal{L}(H\otimes K), \end{align*} where $M\odot N$ denotes the algebraic tensor product represented on $H\otimes K$. [/definition] The spatial tensor product is the operator-topology closure appropriate for von Neumann algebras. The natural question is whether combining two factors can create a new central operator. The obstruction to a naive algebraic argument is that an element of $M\overline{\otimes}N$ need not be a finite sum of elementary tensors, so centrality cannot be checked only coefficient by coefficient in a finite expansion. The center formula answers this by using von Neumann tensor-product tools to reduce the center of the tensor product to the centers of the two inputs. [quotetheorem:9289] [citeproof:9289] This result makes factors stable under spatial tensor product. The spatial von Neumann closure is essential: for algebraic tensor products or incomplete operator-algebra tensor norms, the commutant and slice-map arguments need not apply in this form. The theorem also does not say that every factor decomposes as a tensor product, nor that such decompositions are unique when they exist; tensor-prime phenomena are a separate and subtler subject. Its role here is constructive: it gives a reliable mechanism for building larger factors from smaller ones, such as matrix amplifications and tensor powers. [example: Matrix Amplification of a Factor] Let $M$ be a factor and use the standard identification \begin{align*} M_n(M)=M\overline{\otimes}M_n(\mathbb C). \end{align*} Because $M$ is a factor, its center is \begin{align*} Z(M)=\mathbb C I_M. \end{align*} From the matrix-algebra center computation, \begin{align*} Z(M_n(\mathbb C))=\mathbb C I_n. \end{align*} Applying the *Tensor Product Center Formula* gives \begin{align*} Z(M_n(M))=Z(M\overline{\otimes}M_n(\mathbb C))=Z(M)\overline{\otimes}Z(M_n(\mathbb C)). \end{align*} Substituting the two center computations gives \begin{align*} Z(M_n(M))=(\mathbb C I_M)\overline{\otimes}(\mathbb C I_n). \end{align*} The right-hand side consists exactly of scalar multiples of $I_M\otimes I_n$, since every elementary tensor has the form \begin{align*} (\lambda I_M)\otimes(\mu I_n)=\lambda\mu(I_M\otimes I_n) \end{align*} with $\lambda,\mu\in\mathbb C$, and the weak closure of a one-dimensional subspace is itself. Hence \begin{align*} Z(M_n(M))=\mathbb C(I_M\otimes I_n)=\mathbb C I_{M_n(M)}. \end{align*} Thus $M_n(M)$ has scalar center, so matrix amplification preserves the factor property. [/example] The chapter’s conclusion is that the center is the algebraic shadow of decomposition. Central projections split a von Neumann algebra into sectors; central carriers describe which sectors a projection occupies; factors are precisely the algebras for which this central splitting has disappeared. The next stage of the course keeps the factor assumption and studies how projections inside a factor can still differ by size, equivalence, and trace. When the center is trivial, the algebra is a factor, but even then projections still come in distinct sizes and equivalence classes. The next chapter keeps the factor assumption and shows how minimal projections and matrix units recover the type I part as the discrete side of the classification. # 7. Type I Algebras and Decomposition by Minimal Projections This chapter explains how minimal projections detect the discrete part of a von Neumann algebra. In earlier chapters, projections and Murray-von Neumann equivalence gave a comparison theory inside an algebra; here they become a reconstruction tool. The guiding question is: when does a von Neumann algebra look like bounded operators on Hilbert spaces, possibly varying over a measure space? ## Minimal Projections and Atomic Algebras The first problem is to identify the smallest nonzero pieces of a von Neumann algebra. In a measure algebra such as $L^\infty(X)$, projections are characteristic functions of measurable sets, so minimal projections correspond to atoms of the measure space. In a noncommutative algebra, the same idea is encoded by corners $pMp$. [definition: Minimal Projection] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra. A nonzero projection $p \in M$ is minimal in $M$ if the only projections $q \in M$ satisfying $0 \le q \le p$ are $q=0$ and $q=p$. [/definition] The definition is order-theoretic, but computations inside von Neumann algebras are usually made with compressed algebras. The problem is to replace the absence of smaller projections by a test involving all elements of $pMp$; this is what makes minimality usable in later reconstruction arguments. [quotetheorem:9290] [citeproof:9290] This criterion is useful because it turns order-theoretic minimality into an algebraic computation. The hypothesis that $p$ is a projection in $M$ matters: a rank-one projection on the ambient Hilbert space need not belong to the von Neumann algebra being studied, and then it cannot witness minimality inside that algebra. The spectral theorem is also essential to the converse, since the argument detects smaller projections by cutting the spectrum of a self-adjoint element in the corner. The result does not say that $pH$ must be one-dimensional; it says that the algebra $M$ sees no non-scalar operators on $pH$. The next example separates minimality in the ambient Hilbert space from minimality inside the algebra. [example: Minimal Projections in Diagonal and Full Operator Algebras] On $H=\ell^2(\mathbb N)$, let $M=\ell^\infty(\mathbb N)$ act by diagonal multiplication, so an element $a=(a_k)_{k\in\mathbb N}$ acts by $ae_k=a_ke_k$. Let $p_n$ be the projection onto $\mathbb C e_n$. For $a\in M$ and $\xi\in H$, the vector $p_n\xi$ is $(\xi,e_n)e_n$, hence \begin{align*} p_nap_n\xi=p_na((\xi,e_n)e_n)=p_n((\xi,e_n)a_ne_n)=a_n(\xi,e_n)e_n=a_np_n\xi. \end{align*} Thus $p_nap_n=a_np_n$, so $p_nMp_n\subseteq \mathbb C p_n$. The reverse inclusion holds because $p_n=p_n1p_n\in p_nMp_n$, and therefore $p_nMp_n=\mathbb C p_n$. By the *[Minimal Projection Corner Criterion](/theorems/9290)*, $p_n$ is minimal in $M$. The same projection is also minimal in $\mathcal L(H)$. Indeed, if $q\in\mathcal L(H)$ is a projection with $0\le q\le p_n$, then $\operatorname{ran}(q)\subseteq \operatorname{ran}(p_n)=\mathbb C e_n$. Since a projection onto a subspace of the one-dimensional space $\mathbb C e_n$ has range either $\{0\}$ or $\mathbb C e_n$, we get $q=0$ or $q=p_n$. By contrast, let $X=[0,1]$ with [Lebesgue measure](/page/Lebesgue%20Measure) and let $L^\infty(X)$ act by multiplication on $L^2(X)$. Projections in $L^\infty(X)$ are exactly equivalence classes of characteristic functions $\mathbb 1_E$, because $f^2=f=\overline f$ implies $f(x)\in\{0,1\}$ for almost every $x$. If $\mu(E)>0$, choose $t$ with $0<\mu(E\cap[0,t])<\mu(E)$; such a $t$ exists from continuity of the function $t\mapsto\mu(E\cap[0,t])$. Then $\mathbb 1_{E\cap[0,t]}$ is a nonzero projection strictly below $\mathbb 1_E$. Hence no nonzero projection in $L^\infty([0,1])$ is minimal, so the algebra has no minimal projections. [/example] The example shows that an algebra may have many minimal projections or none at all. The next problem is to name the algebras in which minimal projections are not isolated accidents, but occur below every nonzero projection. [definition: Atomic Von Neumann Algebra] A von Neumann algebra $M$ is atomic if every nonzero projection $q \in M$ dominates a nonzero minimal projection $p \in M$ with $p \le q$. [/definition] The definition says that no nonzero part of the algebra is diffuse. It does not require the algebra to be abelian; full matrix algebras and $\mathcal{L}(\ell^2(\mathbb N))$ are atomic in this sense. [example: Matrix Algebras are Atomic] Let $M=M_n(\mathbb C)$ act on $\mathbb C^n$, and let $p\in M$ be a nonzero projection. We first identify the minimal projections. If $\operatorname{rank}(p)=1$ and $q\in M$ is a projection with $0\le q\le p$, then for every $\eta\in\operatorname{ran}(q)$ we have $q\eta=\eta$. Since $p-q\ge0$, \begin{align*} 0\le ((p-q)\eta,\eta)=(p\eta,\eta)-\|\eta\|^2=\|p\eta\|^2-\|\eta\|^2. \end{align*} Because $p$ is an orthogonal projection, $\|p\eta\|\le \|\eta\|$, so equality holds and $p\eta=\eta$. Thus $\operatorname{ran}(q)\subseteq\operatorname{ran}(p)$. The range of $p$ is one-dimensional, so $\operatorname{ran}(q)$ is either $\{0\}$ or $\operatorname{ran}(p)$, and hence $q=0$ or $q=p$. Therefore every rank-one orthogonal projection is minimal. Conversely, if $\operatorname{rank}(p)\ge2$, choose orthonormal vectors $u,v\in\operatorname{ran}(p)$ and let $r$ be the rank-one projection onto $\mathbb C u$, given by $r\eta=(\eta,u)u$. Since $pu=u$, for every $\eta\in\mathbb C^n$, \begin{align*} pr\eta=p((\eta,u)u)=(\eta,u)pu=(\eta,u)u=r\eta. \end{align*} Also, using $p=p^*$ and $pu=u$, \begin{align*} rp\eta=(p\eta,u)u=(\eta,pu)u=(\eta,u)u=r\eta. \end{align*} Thus $pr=rp=r$, so $0\le r\le p$. But $r\ne0$ and $r\ne p$, since $rv=0$ while $pv=v$. Hence $p$ is not minimal. The minimal projections in $M_n(\mathbb C)$ are exactly the rank-one orthogonal projections. Now let $q\in M_n(\mathbb C)$ be any nonzero projection. Its range is a nonzero subspace of $\mathbb C^n$, so choose a unit vector $u\in\operatorname{ran}(q)$ and let $r\eta=(\eta,u)u$. The same computation with $q$ in place of $p$ gives $qr=rq=r$, hence $0\le r\le q$. Since $r$ has rank one, it is minimal. Therefore every nonzero projection in $M_n(\mathbb C)$ dominates a nonzero minimal projection, so $M_n(\mathbb C)$ is atomic. [/example] The matrix example shows the noncommutative form of atomicity, while the diagonal example suggests a purely measure-theoretic model. The next problem is to prove that, in the abelian case, atomicity leaves no structure beyond a set of atoms and bounded scalar functions on that set. [quotetheorem:9291] [citeproof:9291] This theorem explains why atoms in measure theory and minimal projections in operator algebras are the same structural phenomenon. The abelian hypothesis is doing real work: $M_n(\mathbb C)$ is atomic but is not isomorphic to $\ell^\infty(I)$, because its off-diagonal matrix entries link different minimal projections. The atomicity hypothesis is also necessary, since $L^\infty([0,1])$ has many nonzero projections but no minimal ones under Lebesgue measure, so it cannot be reduced to functions on a discrete set. The theorem classifies only the diagonal, measure-theoretic part of the story; in noncommutative algebras, minimal projections can be linked by partial isometries, and those links produce matrix units. ## Matrix Units and Reconstruction The next question is how to rebuild a noncommutative algebra from its minimal pieces. A single minimal projection gives only a scalar corner. A family of equivalent minimal projections, together with chosen partial isometries between them, remembers the full matrix algebra. [definition: System of Matrix Units] Let $I$ be a nonempty index set and let $M$ be a von Neumann algebra. A family $(e_{ij})_{i,j\in I}$ in $M$ is a system of matrix units if \begin{align*} e_{ij}^* &= e_{ji}, & e_{ij}e_{kl} &= \delta_{jk}e_{il} \end{align*} for all $i,j,k,l\in I$, and each $e_{ii}$ is a nonzero projection. [/definition] The diagonal elements $e_{ii}$ are mutually orthogonal projections. The off-diagonal elements are partial isometries carrying $e_{jj}H$ onto $e_{ii}H$, so matrix units encode Murray-von Neumann equivalence at the level of operators. [example: Standard Matrix Units] Let $(\varepsilon_1,\dots,\varepsilon_n)$ be the standard orthonormal basis of $\mathbb C^n$, and define $e_{ij}\varepsilon_m=\delta_{jm}\varepsilon_i$. Then $e_{ii}\varepsilon_i=\varepsilon_i$ and $e_{ii}\varepsilon_m=0$ for $m\ne i$, so $e_{ii}\ne0$ and $e_{ii}^2=e_{ii}=e_{ii}^*$; hence each $e_{ii}$ is a nonzero projection. For the adjoint, compute on basis vectors: \begin{align*} (e_{ij}\varepsilon_m,\varepsilon_r)=\delta_{jm}(\varepsilon_i,\varepsilon_r)=\delta_{jm}\delta_{ir}. \end{align*} Also, \begin{align*} (\varepsilon_m,e_{ji}\varepsilon_r)=(\varepsilon_m,\delta_{ir}\varepsilon_j)=\delta_{ir}\delta_{mj}. \end{align*} The two scalars are equal for all $m,r$, so $e_{ij}^*=e_{ji}$. For multiplication, apply the product to a basis vector $\varepsilon_m$: \begin{align*} e_{ij}e_{kl}\varepsilon_m=e_{ij}(\delta_{lm}\varepsilon_k)=\delta_{lm}\delta_{jk}\varepsilon_i. \end{align*} On the other hand, \begin{align*} \delta_{jk}e_{il}\varepsilon_m=\delta_{jk}\delta_{lm}\varepsilon_i. \end{align*} These agree for every basis vector, so $e_{ij}e_{kl}=\delta_{jk}e_{il}$. Thus the family $(e_{ij})_{1\le i,j\le n}$ satisfies exactly the defining relations for a system of matrix units in $M_n(\mathbb C)$. [/example] The standard example raises the reconstruction problem: if an abstract von Neumann algebra contains a complete system of matrix units, what is the part of the algebra not already described by those matrix units? The answer is that all remaining information sits in one diagonal corner. [quotetheorem:9292] [citeproof:9292] This theorem is the operator-algebraic version of choosing a basis. The condition $\bigvee_i e_{ii}=1$ is necessary: without it the matrix units reconstruct only the corner supported by $\bigvee_i e_{ii}$, leaving an orthogonal summand of $M$ invisible. The theorem also does not say that every system of equivalent projections gives a factor, because the coefficient corner $N$ may carry a nontrivial center or even be noncommutative. Its role in the sequel is to turn a complete family of equivalent minimal projections into an explicit tensor-product model, which is the mechanism behind the [classification of type I factors](/theorems/9293). [example: Reconstructing $\mathcal{L}(K)$] Let $K$ have orthonormal basis $(\xi_i)_{i\in I}$, let $M=\mathcal L(K)$, and define $e_{ij}\eta=(\eta,\xi_j)_K\xi_i$. For diagonal terms, \begin{align*} e_{ii}^2\eta=e_{ii}((\eta,\xi_i)_K\xi_i)=(\eta,\xi_i)_K(\xi_i,\xi_i)_K\xi_i=(\eta,\xi_i)_K\xi_i=e_{ii}\eta. \end{align*} Also, \begin{align*} (e_{ii}\eta,\zeta)_K=((\eta,\xi_i)_K\xi_i,\zeta)_K=(\eta,\xi_i)_K(\xi_i,\zeta)_K=(\eta,(\zeta,\xi_i)_K\xi_i)_K=(\eta,e_{ii}\zeta)_K, \end{align*} so $e_{ii}=e_{ii}^*$, and $e_{ii}\ne0$ because $e_{ii}\xi_i=\xi_i$. For the adjoint relation, for all $\eta,\zeta\in K$, \begin{align*} (e_{ij}\eta,\zeta)_K=((\eta,\xi_j)_K\xi_i,\zeta)_K=(\eta,\xi_j)_K(\xi_i,\zeta)_K=(\eta,(\zeta,\xi_i)_K\xi_j)_K=(\eta,e_{ji}\zeta)_K. \end{align*} Hence $e_{ij}^*=e_{ji}$. For multiplication, \begin{align*} e_{ij}e_{kl}\eta=e_{ij}((\eta,\xi_l)_K\xi_k)=(\eta,\xi_l)_K(\xi_k,\xi_j)_K\xi_i=\delta_{jk}(\eta,\xi_l)_K\xi_i=\delta_{jk}e_{il}\eta. \end{align*} Thus $e_{ij}e_{kl}=\delta_{jk}e_{il}$, so $(e_{ij})_{i,j\in I}$ is a system of matrix units. If $F\subset I$ is finite, then $\sum_{i\in F}e_{ii}$ is the orthogonal projection onto $\operatorname{span}\{\xi_i:i\in F\}$. Since $(\xi_i)_{i\in I}$ is an orthonormal basis, these finite-dimensional projections converge strongly to $1_K$, so $\bigvee_{i\in I}e_{ii}=1$. Now fix $i_0\in I$. For $T\in\mathcal L(K)$ and $\eta\in K$, \begin{align*} e_{i_0i_0}Te_{i_0i_0}\eta=e_{i_0i_0}T((\eta,\xi_{i_0})_K\xi_{i_0})=(\eta,\xi_{i_0})_K(T\xi_{i_0},\xi_{i_0})_K\xi_{i_0}. \end{align*} Therefore \begin{align*} e_{i_0i_0}Te_{i_0i_0}=(T\xi_{i_0},\xi_{i_0})_K e_{i_0i_0}. \end{align*} This proves $e_{i_0i_0}Me_{i_0i_0}\subseteq\mathbb C e_{i_0i_0}$, while the reverse inclusion holds because $e_{i_0i_0}=e_{i_0i_0}1e_{i_0i_0}$. Hence $e_{i_0i_0}Me_{i_0i_0}=\mathbb C e_{i_0i_0}$. By the *[Matrix Unit Reconstruction Theorem](/theorems/9292)*, $M$ is isomorphic to $\mathbb C e_{i_0i_0}\overline{\otimes}\mathcal L(\ell^2(I))$, which is naturally $\mathcal L(\ell^2(I))$. Under the unitary $U:K\to\ell^2(I)$ defined by $U\xi_i=\delta_i$, this is exactly the usual identification $\mathcal L(K)\cong\mathcal L(\ell^2(I))$. [/example] Matrix units also show why minimal projections in a factor tend to come in a single equivalence class. That observation leads to the classification of type I factors. ## Type I Factors The classification problem for factors asks how much structure remains after the center has collapsed to scalars. For type I factors, the answer is: only the Hilbert space dimension. The key hypothesis is the presence of a minimal projection. [definition: Type I Factor] A factor $M$ is type I if it contains a minimal projection. [/definition] This definition is short because the factor condition does most of the work. The next problem is to show that one minimal projection, together with the absence of central decomposition, forces a complete family of matrix units and hence a full operator algebra. [quotetheorem:9293] [citeproof:9293] The theorem gives the exact meaning of type I in the factor case: the algebra is the full algebra of bounded operators on a Hilbert space. The factor hypothesis cannot be dropped, since $M_2(\mathbb C)\oplus M_3(\mathbb C)$ contains minimal projections but has a two-dimensional center and is not a single full operator algebra. The minimal projection hypothesis is equally restrictive: type II and type III factors have no minimal projections, so the matrix-unit reconstruction used in the proof has no starting point. The uniqueness statement records the only remaining invariant in the factor case, and the next examples show how finite and infinite dimensions appear in this language. [example: Finite and Infinite Type I Factors] In $M_n(\mathbb C)$, let $(e_{ij})_{1\le i,j\le n}$ be the standard matrix units. If $A=(a_{rs})$ lies in the center, then $Ae_{ss}=e_{ss}A$ for each $s$. Multiplying on the left by $e_{rr}$ with $r\ne s$ gives \begin{align*} e_{rr}Ae_{ss}=e_{rr}e_{ss}A=0. \end{align*} But $e_{rr}Ae_{ss}=a_{rs}e_{rs}$, so $a_{rs}=0$ for $r\ne s$. Thus $A$ is diagonal. Now $Ae_{rs}=e_{rs}A$ gives \begin{align*} a_{rr}e_{rs}=a_{ss}e_{rs}. \end{align*} Since $e_{rs}\ne0$, we get $a_{rr}=a_{ss}$ for all $r,s$, hence $A=\lambda 1$. Therefore $Z(M_n(\mathbb C))=\mathbb C1$, so $M_n(\mathbb C)$ is a factor. Its rank-one orthogonal projections are minimal, so $M_n(\mathbb C)$ is a type I factor; under the usual action on $\mathbb C^n$, it is $\mathcal L(\mathbb C^n)$. For $H=\ell^2(\mathbb N)$, let $p_k\xi=(\xi,e_k)e_k$. If $q\in\mathcal L(H)$ is a projection with $0\le q\le p_k$, then $\operatorname{ran}(q)\subseteq\operatorname{ran}(p_k)=\mathbb Ce_k$, so $\operatorname{ran}(q)$ is either $\{0\}$ or $\mathbb Ce_k$. Hence $q=0$ or $q=p_k$, and $p_k$ is minimal. The commutant of $\mathcal L(H)$ inside $\mathcal L(H)$ is $\mathbb C1$, so $\mathcal L(\ell^2(\mathbb N))$ is also a type I factor. These two factors are not isomorphic when $n<\infty$. In $M_n(\mathbb C)$, every nonzero family of mutually orthogonal rank-one projections has at most $n$ members, because their ranges are nonzero mutually orthogonal subspaces of the $n$-dimensional space $\mathbb C^n$. The projections $e_{11},\dots,e_{nn}$ show that the maximum is exactly $n$. In $\mathcal L(\ell^2(\mathbb N))$, the projections $p_k$ are mutually orthogonal rank-one projections for all $k\in\mathbb N$, and they are mutually Murray-von Neumann equivalent via the partial isometries $v_{ij}\xi=(\xi,e_j)e_i$. Thus the finite factor has exactly $n$ orthogonal equivalent minimal pieces, while the infinite one has countably many. [/example] The factor assumption is essential. A direct sum such as $M_2(\mathbb C)\oplus M_3(\mathbb C)$ has many minimal projections, but its center records two summands, so it cannot be a single full operator algebra. [remark: Type I Factors and Irreducibility] A concrete von Neumann algebra $M\subseteq \mathcal{L}(H)$ is a factor exactly when $M\cap M'=\mathbb C1$. When $M=\mathcal{L}(K)$ acts on $K$, its commutant is $\mathbb C1$, so it is a factor. More general representations of $\mathcal{L}(K)$ may have larger commutants, but the abstract algebra remains a factor. [/remark] To understand general type I algebras, we must allow the Hilbert space dimension to vary over the spectrum of the center. A non-factor type I algebra cannot be classified by one Hilbert space dimension: the direct sum $M_2(\mathbb C)\oplus M_3(\mathbb C)$ already has two incompatible fibre dimensions, separated by central projections. The center records where each factor-like component lives, so the intrinsic definition must refer to projections seen after cutting by central summands rather than to a single global model. This moves the classification from a single factor to central pieces, and in separable measurable settings to a measurable family of factors. ## General Type I Von Neumann Algebras The final problem is to describe algebras that are type I at every central point but not necessarily factors. The center behaves like an $L^\infty$ algebra, and over each point of the underlying measure space one sees a type I factor. Before using direct integrals, we need an intrinsic definition that does not already assume the classification theorem. [definition: Abelian Projection] Let $M$ be a von Neumann algebra. A projection $p\in M$ is abelian if the corner von Neumann algebra $pMp$ is abelian. [/definition] Minimal projections are abelian because their corners are scalar, but abelian projections need not be minimal. For example, in $L^\infty(X)\overline{\otimes}M_n(\mathbb C)$, a rank-one matrix projection over a non-atomic central region gives an abelian corner isomorphic to $L^\infty$ of that region. This weaker notion is the right one for type I algebras, since a diffuse center may prevent genuine minimal projections even though every fibre is a full matrix algebra. [definition: Type I Von Neumann Algebra] A von Neumann algebra $M$ is type I if every nonzero central projection $z\in Z(M)$ dominates a nonzero abelian projection $p\in M$ with $p\le z$. [/definition] This definition is intrinsic: it does not assume a prior direct integral decomposition and it applies before any measurable model has been chosen. The quantification over nonzero central projections prevents a hidden diffuse or type II/III summand from being ignored. The remaining problem is to isolate the basic building blocks before allowing dimensions to vary; these fixed-dimension pieces are the homogeneous type I algebras. [definition: Homogeneous Type I Algebra] Let $\kappa$ be a Hilbert space dimension. A type I von Neumann algebra $M$ is homogeneous of type $I_\kappa$ if it is isomorphic to \begin{align*} A\ \overline{\otimes}\ \mathcal{L}(K), \end{align*} where $A$ is an abelian von Neumann algebra and $\dim K=\kappa$. [/definition] Homogeneous algebras are the non-factor version of $\mathcal{L}(K)$ with fixed fibre dimension. The abelian algebra $A$ is the center, and the factor part $\mathcal{L}(K)$ supplies the matrix direction. [example: The Model $L^\infty(X)\overline{\otimes}\mathcal{L}(K)$] Let $(X,\mathcal A,\mu)$ be a measure space and let $K$ be a Hilbert space. Put $M=L^\infty(X)\overline{\otimes}\mathcal{L}(K)$ on $L^2(X;K)$, where elementary tensors act by \begin{align*} ((a\otimes T)\xi)(x)=a(x)T(\xi(x)). \end{align*} The multiplication rule is fibrewise: for $a,b\in L^\infty(X)$, $S,T\in\mathcal L(K)$, and $\xi\in L^2(X;K)$, \begin{align*} ((a\otimes T)(b\otimes S)\xi)(x)=a(x)T(b(x)S\xi(x))=a(x)b(x)TS\xi(x)=((ab)\otimes TS)\xi(x). \end{align*} Similarly, \begin{align*} ((a\otimes T)^*\xi)(x)=\overline{a(x)}\,T^*\xi(x)=((\overline a)\otimes T^*)\xi(x). \end{align*} Every $a\otimes 1_K$ is central, because for each elementary tensor $b\otimes T$, \begin{align*} (a\otimes 1_K)(b\otimes T)=(ab)\otimes T=(ba)\otimes T=(b\otimes T)(a\otimes 1_K). \end{align*} Conversely, a central element $Z\in M$ is an essentially bounded measurable field $x\mapsto Z(x)\in\mathcal L(K)$. Since $Z$ commutes with $1\otimes T$ for every $T\in\mathcal L(K)$, for almost every $x$ we have \begin{align*} Z(x)T=TZ(x)\quad\text{for every }T\in\mathcal L(K). \end{align*} The commutant of $\mathcal L(K)$ inside itself is $\mathbb C1_K$, so $Z(x)=a(x)1_K$ for some scalar function $a$. Essential boundedness of $Z$ gives $a\in L^\infty(X)$, hence \begin{align*} Z(M)=L^\infty(X)\otimes \mathbb C1_K. \end{align*} Thus the center records the measurable base $X$, while over each nonzero central region $\mathbb 1_E\otimes 1_K$ the corner is $L^\infty(E)\overline{\otimes}\mathcal L(K)$. When $K=\mathbb C^n$, this is precisely the algebra of essentially bounded measurable $n\times n$ matrix-valued functions, with pointwise matrix multiplication and adjoint. [/example] The model example covers fixed fibre dimension, but a general central decomposition may contain different dimensions on different central supports. Without separability or a standard measure-space framework, the cleanest classification is by central homogeneous summands; direct integral notation then requires extra representation-theoretic hypotheses. The next theorem states the direct integral form in the standard separable setting used in the course. [quotetheorem:9294] [citeproof:9294] This theorem says that, in the separable standard setting, type I algebras are measurable bundles of full operator algebras. The separability and standardness assumptions are not cosmetic: outside this framework, direct integral notation can hide set-theoretic and measurable-field choices, so the safer general statement is the central decomposition into homogeneous type I summands. The theorem does not collapse the algebra to a single $\mathcal{L}(K)$ unless the center is scalar; the function $x\mapsto\dim K_x$ may vary on disjoint central supports. The center chooses the base measure space, while abelian projections and matrix units identify the fibrewise operator algebra. [example: Varying Fibre Dimension] Let $X=X_1\sqcup X_2$ be a measurable disjoint union with $\mu(X_1)>0$ and $\mu(X_2)>0$, and set \begin{align*} M=L^\infty(X_1)\overline{\otimes}M_2(\mathbb C)\ \oplus\ L^\infty(X_2)\overline{\otimes}\mathcal{L}(\ell^2(\mathbb N)). \end{align*} An element of $M$ is a pair $(A_1,A_2)$, and multiplication and adjoints are coordinatewise: \begin{align*} (A_1,A_2)(B_1,B_2)=(A_1B_1,A_2B_2). \end{align*} Thus $(A_1,A_2)$ is central in $M$ exactly when $A_1$ is central in $L^\infty(X_1)\overline{\otimes}M_2(\mathbb C)$ and $A_2$ is central in $L^\infty(X_2)\overline{\otimes}\mathcal{L}(\ell^2(\mathbb N))$. For the first summand, a central element is an essentially bounded measurable matrix field $x\mapsto A_1(x)\in M_2(\mathbb C)$ satisfying \begin{align*} A_1(x)T=TA_1(x) \end{align*} for every $T\in M_2(\mathbb C)$ and almost every $x\in X_1$. Writing $A_1(x)=(a_{rs}(x))_{r,s=1}^2$, commutation with the diagonal matrix unit $e_{11}$ gives $a_{12}(x)=a_{21}(x)=0$, and commutation with $e_{12}$ gives $a_{11}(x)=a_{22}(x)$. Hence $A_1(x)=a_1(x)1_2$ for some $a_1\in L^\infty(X_1)$. The same argument in the second summand uses rank-one operators: if $A_2(x)$ commutes with every rank-one operator $\theta_{\eta,\zeta}\rho=(\rho,\zeta)\eta$ on $\ell^2(\mathbb N)$, then \begin{align*} A_2(x)\theta_{\eta,\zeta}\rho=(\rho,\zeta)A_2(x)\eta \end{align*} and \begin{align*} \theta_{\eta,\zeta}A_2(x)\rho=(A_2(x)\rho,\zeta)\eta. \end{align*} Taking $\rho=\zeta$ with $\|\zeta\|=1$ shows $A_2(x)\eta=(A_2(x)\zeta,\zeta)\eta$ for every $\eta$ orthogonal to no fixed coordinate choice after varying $\eta$, so $A_2(x)$ is a scalar multiple of $1_{\ell^2(\mathbb N)}$. Therefore \begin{align*} Z(M)=\bigl(L^\infty(X_1)\otimes \mathbb C1_2\bigr)\oplus\bigl(L^\infty(X_2)\otimes \mathbb C1_{\ell^2(\mathbb N)}\bigr)\cong L^\infty(X_1)\oplus L^\infty(X_2)\cong L^\infty(X). \end{align*} Now let $z\in Z(M)$ be a nonzero central projection. Under the above identification, $z=(\mathbb 1_E,\mathbb 1_F)$ for measurable sets $E\subseteq X_1$ and $F\subseteq X_2$, with $\mu(E)>0$ or $\mu(F)>0$. If $\mu(E)>0$, let $r$ be the rank-one projection in $M_2(\mathbb C)$ onto $\mathbb C(1,0)$. Then \begin{align*} p=(\mathbb 1_E\otimes r,0) \end{align*} is a nonzero projection with $p\le z$. Its corner is \begin{align*} pMp=(\mathbb 1_E\otimes r)\bigl(L^\infty(X_1)\overline{\otimes}M_2(\mathbb C)\bigr)(\mathbb 1_E\otimes r)\oplus 0. \end{align*} For $a\in L^\infty(X_1)$ and $T\in M_2(\mathbb C)$, \begin{align*} (\mathbb 1_E\otimes r)(a\otimes T)(\mathbb 1_E\otimes r)=(\mathbb 1_Ea\mathbb 1_E)\otimes rTr. \end{align*} Since $rTr=(Te_1,e_1)r$, this corner is isomorphic to $L^\infty(E)$ and is abelian. If instead $\mu(F)>0$, choose the rank-one projection $s\xi=(\xi,e_1)e_1$ on $\ell^2(\mathbb N)$ and set \begin{align*} p=(0,\mathbb 1_F\otimes s). \end{align*} For $T\in\mathcal L(\ell^2(\mathbb N))$ and $\xi\in\ell^2(\mathbb N)$, \begin{align*} sTs\xi=sT((\xi,e_1)e_1)=(\xi,e_1)(Te_1,e_1)e_1=(Te_1,e_1)s\xi, \end{align*} so $s\mathcal L(\ell^2(\mathbb N))s=\mathbb Cs$, and the corner $pMp$ is isomorphic to $L^\infty(F)$. Thus every nonzero central projection dominates a nonzero abelian projection, so $M$ is type I. The fibre over $X_1$ is $M_2(\mathbb C)\cong\mathcal L(\mathbb C^2)$, so its Hilbert space dimension is $2$. The fibre over $X_2$ is $\mathcal L(\ell^2(\mathbb N))$, so its Hilbert space dimension is countably infinite. The central projection $(\mathbb 1_{X_1},0)$ separates the finite-dimensional part from the countably infinite part. [/example] The classification also clarifies what is special about atomic type I algebras. If the center is atomic, the direct integral collapses to an $\ell^\infty$-direct product of type I factors; if the center is diffuse, the same fibrewise factors are spread over a continuous measure space. [remark: Atomic Versus Diffuse Center] A type I algebra can have minimal projections only over atoms of its center. For instance, $L^\infty([0,1])\overline{\otimes}M_n(\mathbb C)$ has no minimal projections because no nonzero central projection in $L^\infty([0,1])$ is minimal. In contrast, $\ell^\infty(I)\overline{\otimes}M_n(\mathbb C)$ is atomic and decomposes as a bounded product of copies of $M_n(\mathbb C)$. [/remark] The chapter's main picture is therefore a hierarchy. Minimal projections give scalar corners; matrix units glue equivalent minimal projections into full operator algebras; factors remove the center and become $\mathcal{L}(K)$; general type I algebras reintroduce the center as a measurable base over which these factors vary. Type I algebras show how minimal projections reconstruct familiar matrix behavior, yet the comparison theory developed earlier suggests that not all factors are discrete. The next chapter moves beyond minimal projections to traces and semifinite dimension, where Type II algebras sit between finite matrix-like structure and the infinite world. # 8. Type II Algebras and Traces This chapter turns the type classification from projection comparison into numerical structure. It assumes the preceding material on von Neumann algebras, factors, centres, projections, Murray-von Neumann equivalence, partial isometries, finite and properly infinite projections, and normal positive functionals. In a Type I factor, the rank of a projection is an integer or infinity; in a Type II algebra, projections still have sizes, but those sizes vary continuously. The tool that measures them is a trace, first in the finite normalized setting and then in the semifinite setting needed for Type $II_\infty$ algebras. ## Tracial States and Semifinite Traces The projection comparison theory from Chapter 5 and the type I classification from Chapter 7 tell us when many projections are Murray-von Neumann equivalent, but they do not yet assign a numerical size to an equivalence class. The guiding question is: when can projection comparison be represented by an additive, normal, unitary-invariant functional? A state on a von Neumann algebra is useful for dimension theory only if it respects the operator-algebraic closure and is insensitive to cyclic permutations. This leads to the finite trace used throughout Type $II_1$ theory. [definition: Faithful Normal Tracial State] Let $M$ be a von Neumann algebra. A faithful normal tracial state on $M$ is a map $\tau:M\to \mathbb C$ such that: 1. $\tau$ is linear; 2. $\tau(x^*x)\ge 0$ for all $x\in M$; 3. $\tau(1)=1$; 4. $\tau(x^*x)=0$ implies $x=0$; 5. for every bounded increasing net $(a_i)$ of positive elements of $M$ with supremum $a$, $\tau(a_i)\uparrow \tau(a)$; 6. $\tau(xy)=\tau(yx)$ for all $x,y\in M$. [/definition] The normality condition is the bridge from algebra to von Neumann algebra: it makes the trace compatible with suprema of projections and with weak operator limits. Faithfulness prevents a nonzero projection from having zero size, while the tracial identity makes the value depend only on Murray-von Neumann equivalence. As in Chapter 6, $Z(M)$ denotes the center of a von Neumann algebra $M$, and $M$ is a factor precisely when $Z(M)=\mathbb C1$. [example: The Matrix Trace] For $M=M_n(\mathbb C)$, define $\tau(a)=n^{-1}\operatorname{Tr}(a)$. [Linearity and positivity](/theorems/4904) follow from the ordinary matrix trace: if $a=x^*x$, then \begin{align*} \operatorname{Tr}(x^*x)=\sum_{i,j=1}^n |x_{ij}|^2\ge 0. \end{align*} Also $\tau(1)=n^{-1}\operatorname{Tr}(I_n)=n^{-1}n=1$, and if $\tau(x^*x)=0$, then $\sum_{i,j}|x_{ij}|^2=0$, so every entry of $x$ is $0$. Since $M_n(\mathbb C)$ is finite-dimensional, every increasing bounded net of positive matrices has entrywise, norm, and weak operator limits agreeing, and the ordinary trace is continuous for this limit; hence $\tau$ is normal. Finally, for matrices $a=(a_{ij})$ and $b=(b_{ij})$, \begin{align*} \operatorname{Tr}(ab)=\sum_{i=1}^n\sum_{j=1}^n a_{ij}b_{ji}. \end{align*} Renaming the indices gives \begin{align*} \sum_{i=1}^n\sum_{j=1}^n a_{ij}b_{ji}=\sum_{j=1}^n\sum_{i=1}^n b_{ji}a_{ij}=\operatorname{Tr}(ba), \end{align*} so $\tau(ab)=\tau(ba)$. Now let $p\in M_n(\mathbb C)$ be a projection. Since $p=p^*=p^2$, its eigenvalues satisfy $\lambda=\lambda^2$, so each eigenvalue is $0$ or $1$. The ordinary trace is the sum of the eigenvalues counted with multiplicity, while $\operatorname{rank}(p)$ is the multiplicity of the eigenvalue $1$. Therefore \begin{align*} \tau(p)=\frac{1}{n}\operatorname{Tr}(p)=\frac{\operatorname{rank}(p)}{n}. \end{align*} Because $\operatorname{rank}(p)$ can only be one of $0,1,\dots,n$, the possible trace values are exactly \begin{align*} \{0,1/n,\dots,1\}. \end{align*} Thus the normalized matrix trace is precisely normalized dimension for projections in $M_n(\mathbb C)$. [/example] The matrix example has only one possible normalized trace, because all minimal projections are equivalent and add up to the identity. The same uniqueness question matters in factors: if a Type $II_1$ factor had two normalized traces, the phrase "the dimension of a projection" would be ambiguous. The next theorem removes that ambiguity by using the absence of a nontrivial centre. [quotetheorem:9295] [citeproof:9295] The factorial hypothesis is doing real work here. If $M=M_n(\mathbb C)\oplus M_m(\mathbb C)$, then convex combinations of the two normalized matrix traces give many normalized tracial states, because the nontrivial centre records how much weight is assigned to each summand. Faithfulness and normality also matter: without faithfulness, a trace may ignore a nonzero central summand, and without normality it need not interact correctly with projection suprema. The theorem does not say that every factor has a trace; Type $III$ factors have no nonzero semifinite normal trace, and Type $I_\infty$ factors have no tracial state. What it gives is uniqueness once the finite tracial situation exists, which justifies writing $\tau$ without an extra choice. The next enlargement is needed because Type $II_\infty$ algebras have no finite normalized trace on the identity, but their finite corners still carry finite trace. [definition: Faithful Normal Semifinite Trace] Let $M$ be a von Neumann algebra. A faithful normal semifinite trace on $M$ is a map $\operatorname{Tr}:M_+\to [0,\infty]$ such that: 1. $\operatorname{Tr}(a+b)=\operatorname{Tr}(a)+\operatorname{Tr}(b)$ for all $a,b\in M_+$; 2. $\operatorname{Tr}(\lambda a)=\lambda\operatorname{Tr}(a)$ for all $a\in M_+$ and $\lambda\ge 0$; 3. $\operatorname{Tr}(x^*x)=\operatorname{Tr}(xx^*)$ for all $x\in M$; 4. $\operatorname{Tr}(a)=0$ implies $a=0$ for $a\in M_+$; 5. for every bounded increasing net $(a_i)$ in $M_+$ with supremum $a$, $\operatorname{Tr}(a_i)\uparrow \operatorname{Tr}(a)$; 6. for every nonzero $a\in M_+$, there exists $0\le b\le a$ with $0<\operatorname{Tr}(b)<\infty$. [/definition] Semifiniteness says that infinite mass can be detected through finite pieces. It is the trace-theoretic analogue of measuring an infinite measure space by finite-measure subsets rather than by a probability measure. [example: The Standard Semifinite Trace on $B(\ell^2)$] Let $(e_n)_{n\in\mathbb N}$ be the standard orthonormal basis of $\ell^2$ and let $M=B(\ell^2)$. For $a\in M_+$ define \begin{align*} \operatorname{Tr}(a)=\sum_{n=1}^{\infty}(ae_n,e_n)_{\ell^2}. \end{align*} Each summand is nonnegative because $a\ge 0$, so the series is well-defined in $[0,\infty]$. If $a,b\in M_+$ and $\lambda\ge 0$, then for each $n$, \begin{align*} ((a+b)e_n,e_n)=(ae_n,e_n)+(be_n,e_n) \end{align*} and \begin{align*} ((\lambda a)e_n,e_n)=\lambda(ae_n,e_n). \end{align*} Summing these nonnegative identities over $n$ gives $\operatorname{Tr}(a+b)=\operatorname{Tr}(a)+\operatorname{Tr}(b)$ and $\operatorname{Tr}(\lambda a)=\lambda\operatorname{Tr}(a)$. For $x\in B(\ell^2)$, write $x_{mn}=(xe_n,e_m)$. Then \begin{align*} (x^*xe_n,e_n)=\|xe_n\|^2=\sum_{m=1}^{\infty}|x_{mn}|^2. \end{align*} Also \begin{align*} (xx^*e_m,e_m)=\|x^*e_m\|^2=\sum_{n=1}^{\infty}|x_{mn}|^2. \end{align*} Therefore both $\operatorname{Tr}(x^*x)$ and $\operatorname{Tr}(xx^*)$ are the same nonnegative double series $\sum_{m,n}|x_{mn}|^2$, so $\operatorname{Tr}(x^*x)=\operatorname{Tr}(xx^*)$. The trace is faithful. If $a\in M_+$ and $\operatorname{Tr}(a)=0$, then every nonnegative summand $(ae_n,e_n)$ is $0$. Since $a^{1/2}$ exists and \begin{align*} (ae_n,e_n)=\|a^{1/2}e_n\|^2, \end{align*} we get $a^{1/2}e_n=0$ for every $n$. By linearity and density of the span of the basis, $a^{1/2}=0$, hence $a=0$. For normality, let $(a_i)$ be a bounded increasing net in $M_+$ with supremum $a$. For each fixed $n$, the numbers $(a_i e_n,e_n)$ increase to $(ae_n,e_n)$. Hence \begin{align*} \sup_i\operatorname{Tr}(a_i)=\sup_i\sum_{n=1}^{\infty}(a_i e_n,e_n). \end{align*} Because all terms are nonnegative and increasing in $i$, this equals \begin{align*} \sup_N\sum_{n=1}^{N}(ae_n,e_n). \end{align*} The last expression is exactly $\sum_{n=1}^{\infty}(ae_n,e_n)=\operatorname{Tr}(a)$, so $\operatorname{Tr}(a_i)\uparrow\operatorname{Tr}(a)$. Finally, the trace is semifinite. If $0\ne a\in M_+$, choose $\xi\in\ell^2$ with $\|\xi\|=1$ and $(a\xi,\xi)>0$. Since $a^{1/2}\xi\ne 0$, define the rank-one projection $r$ onto the line spanned by $a^{1/2}\xi$. The positive operator $b=a^{1/2}ra^{1/2}$ satisfies $0\le b\le a$ because $0\le r\le 1$. It is nonzero, and its range lies in the one-dimensional span of $a^{1/2}r\ell^2$, so it has finite rank. For a positive rank-one operator $b=\alpha(\,\cdot\,,\eta)\eta$ with $\|\eta\|=1$ and $\alpha>0$, \begin{align*} \operatorname{Tr}(b)=\sum_{n=1}^{\infty}\alpha |(e_n,\eta)|^2=\alpha\|\eta\|^2=\alpha. \end{align*} Thus $0<\operatorname{Tr}(b)<\infty$. Since $(1e_n,e_n)=1$ for every $n$, \begin{align*} \operatorname{Tr}(1)=\sum_{n=1}^{\infty}1=\infty. \end{align*} If $p$ is a finite-rank projection of rank $k$, choose an orthonormal basis so that $p$ is the identity on its $k$-dimensional range and zero on its orthogonal complement. In that basis the trace sum has exactly $k$ terms equal to $1$ and all remaining terms equal to $0$, so $\operatorname{Tr}(p)=k$. Thus this trace extends ordinary finite-dimensional dimension while assigning infinite size to the identity of $B(\ell^2)$. [/example] ## Type $II_1$ Factors and Continuous Projection Dimension The central phenomenon in Type $II_1$ theory is that finite algebras need not have discrete projection sizes. The question is: what replaces the integer-valued rank function when a factor is finite but has no minimal projections? [definition: Type $II_1$ Factor] A Type $II_1$ factor is a factor $M$ such that: 1. $M$ is finite; 2. $M$ is infinite-dimensional; 3. $M$ has no nonzero minimal projections; 4. $M$ admits a faithful normal tracial state. [/definition] This definition isolates the finite continuous case in the type classification: finite means the identity has normalized size, while the absence of minimal projections prevents the projection scale from stopping at atoms. To use this class in computations, the abstract trace must be turned into a named projection invariant. The next definition records that invariant so comparison statements can be written numerically. [definition: Dimension Function of a Type $II_1$ Factor] Let $M$ be a Type $II_1$ factor with faithful normal tracial state $\tau$. The dimension function on projections of $M$ is the map \begin{align*} d_M:\mathcal P(M)&\to [0,1], & d_M(p)&=\tau(p). \end{align*} [/definition] The tracial identity implies $d_M(p)=d_M(q)$ whenever $p\sim q$ in the Murray-von Neumann sense. Additivity on orthogonal projections turns decompositions of $1$ into decompositions of the interval $[0,1]$. [example: Matrix Subalgebras Inside a Type $II_1$ Factor] Suppose $M$ contains a unital copy of $M_n(\mathbb C)$ with matrix units $(e_{ij})_{1\le i,j\le n}$, so $e_{ij}e_{kl}=\delta_{jk}e_{il}$, $e_{ij}^*=e_{ji}$, and $\sum_{i=1}^n e_{ii}=1$. For each $i$, the element $e_{i1}$ is a partial isometry with \begin{align*} e_{i1}^*e_{i1}=e_{1i}e_{i1}=e_{11}. \end{align*} Also \begin{align*} e_{i1}e_{i1}^*=e_{i1}e_{1i}=e_{ii}. \end{align*} Thus $e_{11}\sim e_{ii}$ for every $i$. Using the tracial identity, \begin{align*} \tau(e_{ii})=\tau(e_{i1}e_{1i})=\tau(e_{1i}e_{i1})=\tau(e_{11}). \end{align*} Since the diagonal matrix units are orthogonal projections and sum to $1$, \begin{align*} 1=\tau(1)=\tau\left(\sum_{i=1}^n e_{ii}\right)=\sum_{i=1}^n \tau(e_{ii})=\sum_{i=1}^n \tau(e_{11})=n\tau(e_{11}). \end{align*} Therefore $\tau(e_{11})=1/n$, and hence $\tau(e_{ii})=1/n$ for every $i$. For $1\le k\le n$, the projection $p=e_{11}+\cdots+e_{kk}$ is a sum of $k$ pairwise orthogonal projections, so additivity of the trace gives \begin{align*} d_M(p)=\tau(p)=\sum_{i=1}^k \tau(e_{ii})=\sum_{i=1}^k \frac{1}{n}=\frac{k}{n}. \end{align*} Thus the dimension function on this finite matrix block agrees with normalized matrix rank, and the block contributes the rational dimensions $0,1/n,\dots,1$ inside the Type $II_1$ factor. [/example] The example explains why rational dimensions appear whenever matrix units are present, but Type $II_1$ factors have no smallest nonzero projection that would force discreteness. Repeated subdivision and normality suggest that the trace range should have no gaps. The resulting theorem is one of the signatures of Type $II_1$ factors: their projection dimension is continuous rather than discrete. [quotetheorem:9296] This theorem is usually established after developing the projection-comparison machinery for finite factors: one repeatedly cuts nonzero projections into smaller equivalent pieces, uses maximal families of orthogonal subprojections with trace bounded by the target value, and applies normality to take the supremum of the resulting increasing family. In this course the theorem is used as a structural result rather than proved at this point; a full proof belongs with the later projection-lattice arguments that establish comparability and refinement of projections in finite factors. The statement says that every real number between $0$ and $1$ occurs as the dimension of a closed subspace represented internally by a projection in $M$. Each hypothesis is doing work. Factoriality removes central summands; without it, a finite direct sum such as $M_2(\mathbb C)\oplus M_3(\mathbb C)$ has projection data split across two independent centres rather than a single continuous dimension scale. Finiteness gives a normalized trace with $\tau(1)=1$; in a properly infinite algebra the identity may be equivalent to a proper subprojection, so finite-dimensional rank intuition breaks down. The absence of minimal projections prevents atomic stopping points: in $M_n(\mathbb C)$ every nonzero projection contains a rank-one subprojection, and normalized traces are forced into the discrete set $\{0,1/n,\dots,1\}$. The theorem is an existence result, not a construction formula: different procedures may produce non-identical projections with the same trace, and finite-factor comparison theory is what identifies those projections as Murray-von Neumann equivalent. Its role here is to replace the discrete Hilbert-space dimension scale by a continuous trace scale. [example: Continuous Dimensions] Let $M$ be a Type $II_1$ factor with faithful normal tracial state $\tau$, and choose \begin{align*} t=\frac{\sqrt{2}}{2}. \end{align*} Since $0<\sqrt{2}<2$, dividing by $2$ gives $0<t<1$. By the *Dimension Range Theorem for Type $II_1$ Factors*, there is a projection $p\in M$ such that \begin{align*} \tau(p)=t=\frac{\sqrt{2}}{2}. \end{align*} This trace value cannot occur as the normalized rank of a projection in any finite-dimensional matrix algebra. Indeed, if $q\in M_n(\mathbb C)$ is a projection, then the matrix-trace computation gives \begin{align*} \tau_n(q)=\frac{1}{n}\operatorname{Tr}(q)=\frac{\operatorname{rank}(q)}{n}. \end{align*} Writing $k=\operatorname{rank}(q)$, this value is $k/n$, hence rational. On the other hand, $\sqrt{2}/2$ is irrational: if $\sqrt{2}/2=a/b$ with integers $a,b$ and $b\ne 0$, then multiplying by $2$ gives $\sqrt{2}=2a/b$, so $\sqrt{2}$ would be rational, contradicting the irrationality of $\sqrt{2}$. Thus the projection $p$ has a dimension value that is impossible in finite-dimensional matrix algebras, isolating a genuinely Type $II_1$ phenomenon. [/example] Continuous range supplies many projections, but comparison theory asks for more: it asks whether trace is a complete test for subequivalence. In finite-dimensional spaces, $\dim E\le\dim F$ exactly when $E$ embeds isometrically into $F$ after choosing subspaces. The next theorem is the von Neumann algebra version, with partial isometries replacing linear inclusions of subspaces. [quotetheorem:9297] [citeproof:9297] The [equivalence relation](/page/Equivalence%20Relation) itself is then detected by equality of trace values: $p\sim q$ precisely when both inequalities hold. Factoriality is essential because in a direct sum of finite factors, projection comparison can differ on different central summands, so a single total order by trace need not describe subequivalence. Finiteness is also essential: in a properly infinite factor, a projection can be equivalent to a proper subprojection of itself, so trace-like size no longer behaves like finite-dimensional rank. The theorem is not a recipe for constructing the partial isometry; it says that once finite-factor comparison theory is available, the numerical trace is a complete test for whether the required partial isometry exists. [example: Equality of Trace Gives Equivalence] Let $M$ be a finite factor with faithful normal tracial state $\tau$, and let $p,q\in M$ be projections with $\tau(p)=\tau(q)$. Since $\tau(p)\le \tau(q)$, *Trace Comparison of Projections in a Finite Factor* gives $p\precsim q$. Thus there is a projection $r\le q$ such that $p\sim r$. Choose a partial isometry $v\in M$ with $v^*v=p$ and $vv^*=r$. By the tracial identity, \begin{align*} \tau(p)=\tau(v^*v)=\tau(vv^*)=\tau(r). \end{align*} Since $r\le q$, the operator $q-r$ is a projection and $q=r+(q-r)$ with $r(q-r)=0$. Additivity of $\tau$ on orthogonal positive elements gives \begin{align*} \tau(q)=\tau(r)+\tau(q-r). \end{align*} Using $\tau(r)=\tau(p)$ and $\tau(p)=\tau(q)$, this becomes \begin{align*} \tau(q)=\tau(q)+\tau(q-r). \end{align*} Subtracting $\tau(q)$ gives $\tau(q-r)=0$. Faithfulness of $\tau$ forces $q-r=0$, so $q=r$. Hence $p\sim r=q$, and therefore $p\sim q$. Thus, in a finite factor, equality of trace is already strong enough to force Murray-von Neumann equivalence of projections, just as equal dimension forces isomorphism of finite-dimensional subspaces. [/example] ## Hyperfinite and Non-Hyperfinite Examples The abstract definition of a Type $II_1$ factor would be hard to trust without examples. The next problem is to see both a highly approximable factor and a factor that cannot be built from finite-dimensional algebras in that way. [definition: Hyperfinite $II_1$ Factor] A Type $II_1$ factor $R$ is hyperfinite if there exists an increasing sequence of finite-dimensional unital $*$-subalgebras \begin{align*} A_1\subset A_2\subset\cdots\subset R \end{align*} such that $R$ is the weak operator closure of $\bigcup_{n=1}^{\infty}A_n$. [/definition] Hyperfiniteness is a von Neumann algebra analogue of approximation by matrices, with weak closure replacing norm closure. It is one of the reasons Type $II_1$ theory sits between finite-dimensional linear algebra and infinite-dimensional dynamics. Amenability enters as the dynamical condition that finite pieces can almost model the whole action: for amenable groups and amenable measured equivalence relations, this finite approximation is reflected on the von Neumann algebra side by hyperfiniteness. Traces then behave like invariant probability measures, while projections behave like measurable pieces with continuous measure. Nonamenability is not merely a group-theoretic label here; it supplies obstructions to approximating the associated factor by finite-dimensional subalgebras in the weak operator topology. [example: The Hyperfinite $II_1$ Factor] Let $A_n=M_{2^n}(\mathbb C)$, and embed $A_n$ into $A_{n+1}$ by $\iota_n(a)=a\otimes I_2$. If $\tau_n=2^{-n}\operatorname{Tr}_{2^n}$ is the normalized matrix trace, then the traces are compatible under these inclusions. Indeed, for $a\in M_{2^n}(\mathbb C)$, \begin{align*} \tau_{n+1}(a\otimes I_2)=\frac{1}{2^{n+1}}\operatorname{Tr}_{2^{n+1}}(a\otimes I_2). \end{align*} The ordinary trace of a tensor product satisfies $\operatorname{Tr}(a\otimes I_2)=\operatorname{Tr}(a)\operatorname{Tr}(I_2)$, so \begin{align*} \tau_{n+1}(a\otimes I_2)=\frac{1}{2^{n+1}}\operatorname{Tr}_{2^n}(a)\cdot 2. \end{align*} Hence \begin{align*} \tau_{n+1}(a\otimes I_2)=\frac{1}{2^n}\operatorname{Tr}_{2^n}(a)=\tau_n(a). \end{align*} Using these compatible traces, form the GNS representation of the increasing union $\bigcup_{n=1}^{\infty}A_n$, and let $R$ be its weak operator closure. By construction, \begin{align*} A_1\subset A_2\subset\cdots\subset R \end{align*} and $R$ is the weak operator closure of their union, so $R$ is the hyperfinite Type $II_1$ factor. The trace $\tau$ on $R$ restricts to $\tau_n$ on each finite stage $A_n$. If $p\in A_n$ is a projection of matrix rank $k$, then the matrix-trace computation gives \begin{align*} \tau(p)=\tau_n(p)=\frac{1}{2^n}\operatorname{Tr}_{2^n}(p)=\frac{k}{2^n}. \end{align*} Thus every projection visible at the finite stage $A_n$ has dyadic rational dimension. The continuous projection scale of $R$ is not contained in any one finite matrix stage: applying the *Dimension Range Theorem for Type $II_1$ Factors* to $R$ gives projections of every trace value in $[0,1]$. [/example] The previous example is maximally approximable by finite-dimensional algebras. Group von Neumann algebras provide a contrasting family where the group structure can obstruct hyperfiniteness. [example: The Free Group Factor $L(F_2)$] Let $F_2=\langle s,t\rangle$ and let $\lambda:F_2\to B(\ell^2(F_2))$ be the left regular representation, so $\lambda_g\delta_h=\delta_{gh}$. The group von Neumann algebra is \begin{align*} L(F_2)=\lambda(F_2)''. \end{align*} For a finite linear combination $x=\sum_{g\in E}a_g\lambda_g$, where $E\subset F_2$ is finite, the canonical vector trace satisfies \begin{align*} \tau(x)=(x\delta_e,\delta_e)=\left(\sum_{g\in E}a_g\delta_g,\delta_e\right)=a_e. \end{align*} This functional is tracial on the group algebra. If $x=\sum_{g\in E}a_g\lambda_g$ and $y=\sum_{h\in F}b_h\lambda_h$, then \begin{align*} xy=\sum_{g\in E}\sum_{h\in F}a_gb_h\lambda_{gh}. \end{align*} The coefficient of $\lambda_e$ in $xy$ comes from pairs with $gh=e$, equivalently $h=g^{-1}$, so \begin{align*} \tau(xy)=\sum_{g\in E\cap F^{-1}}a_gb_{g^{-1}}. \end{align*} Similarly, the coefficient of $\lambda_e$ in $yx$ comes from pairs with $hg=e$, equivalently $g=h^{-1}$, so \begin{align*} \tau(yx)=\sum_{h\in F\cap E^{-1}}b_ha_{h^{-1}}. \end{align*} Replacing $h$ by $g^{-1}$ in the second sum gives the same terms as the first sum, and complex scalars commute, so $\tau(xy)=\tau(yx)$. The trace is positive and faithful on the group algebra. Since $\lambda_g^*=\lambda_{g^{-1}}$, we have \begin{align*} x^*=\sum_{g\in E}\overline{a_g}\lambda_{g^{-1}}. \end{align*} Therefore \begin{align*} x^*x=\sum_{g\in E}\sum_{h\in E}\overline{a_g}a_h\lambda_{g^{-1}h}. \end{align*} The coefficient of $\lambda_e$ occurs exactly when $g^{-1}h=e$, equivalently $h=g$, so \begin{align*} \tau(x^*x)=\sum_{g\in E}|a_g|^2. \end{align*} Thus $\tau(x^*x)\ge 0$, and $\tau(x^*x)=0$ forces every $a_g=0$, hence $x=0$. The vector functional extends normally from the group algebra to $L(F_2)$, giving the canonical faithful normal tracial state. The free group $F_2$ is ICC: every non-identity element has an infinite conjugacy class. By the *ICC Group [Factor Theorem](/theorems/3235)*, $L(F_2)$ is therefore a factor with its canonical faithful normal tracial state. It is infinite-dimensional because the unitaries $\lambda_g$ are linearly independent for distinct $g\in F_2$: if $\sum_{g\in E}a_g\lambda_g=0$, then applying the operator to $\delta_e$ gives $\sum_{g\in E}a_g\delta_g=0$, and orthonormality of the basis $(\delta_g)_{g\in F_2}$ forces every $a_g=0$. Hence $L(F_2)$ is a Type $II_1$ factor. Since $F_2$ is nonamenable, the *Murray-von Neumann Non-Hyperfiniteness Theorem for Free Group Factors* gives that $L(F_2)$ is not hyperfinite. Thus $L(F_2)$ is a Type $II_1$ factor whose trace is explicit on group elements, but whose nonamenable group origin prevents it from being built as an increasing weak operator closure of finite-dimensional matrix algebras. [/example] The contrast between $R$ and $L(F_2)$ motivates classification questions beyond the type system. Type records the projection scale and trace behaviour, while finer invariants distinguish factors of the same type. From the group-dynamical viewpoint, $L(F_2)$ packages noncommutative information from the left regular action of a nonamenable group, so its failure of hyperfiniteness is not an accident of presentation but a signal that the underlying dynamics cannot be approximated by finite matrix pieces in the amenable way. ## Type $II_\infty$ Algebras and Amplification Finite trace is too restrictive for many naturally infinite algebras built from Type $II_1$ factors. The final question of the chapter is how to pass from a finite continuous dimension scale $[0,1]$ to an infinite continuous scale $[0,\infty]$. [definition: Type $II_\infty$ Factor] A Type $II_\infty$ factor is a factor $N$ such that: 1. $N$ admits a faithful normal semifinite trace; 2. $N$ has no nonzero minimal projections; 3. the identity $1_N$ is an infinite projection. [/definition] For such a factor, the faithful normal semifinite trace assigns infinite trace to the identity. Hence there is no faithful normal tracial state on $N$, while finite corners may still carry normalized traces. A Type $II_\infty$ factor has finite projections of every finite trace size, but its identity has infinite trace. To construct such algebras from the finite theory, we need an operation that keeps the Type $II_1$ factor in one tensor component and supplies countably infinite matrix size in the other. That operation is amplification. [definition: Amplification] Let $M$ be a von Neumann algebra and let $K$ be a Hilbert space. The amplification of $M$ by $K$ is the von Neumann algebra \begin{align*} M\overline{\otimes}B(K), \end{align*} acting on the Hilbert space tensor product associated to a faithful normal representation of $M$. [/definition] For Type $II_1$ factors, amplification by $B(\ell^2)$ turns the normalized trace into a semifinite trace. This produces the model Type $II_\infty$ object attached to a finite factor. [example: The $II_\infty$ Amplification] Let $M$ be a Type $II_1$ factor with faithful normal tracial state $\tau$, let $(e_n)_{n\in\mathbb N}$ be the standard basis of $\ell^2$, and set \begin{align*} N=M\overline{\otimes}B(\ell^2). \end{align*} Let $\operatorname{Tr}$ denote the standard semifinite trace on $B(\ell^2)$, so $\operatorname{Tr}(e_{11})=1$ and $\operatorname{Tr}(1)=\sum_{n=1}^{\infty}(e_n,e_n)=\sum_{n=1}^{\infty}1=\infty$. On positive elementary tensors define \begin{align*} T(a\otimes b)=\tau(a)\operatorname{Tr}(b), \end{align*} and extend normally to $N_+$; this is the tensor-product trace $T=\tau\otimes\operatorname{Tr}$. For positive elementary tensors $a\otimes b$ and $c\otimes d$, additivity and homogeneity follow from the two traces: \begin{align*} T((a\otimes b)+(c\otimes d))=T(a\otimes b)+T(c\otimes d) \end{align*} whenever the tensors are combined inside the positive cone by the [normal extension](/page/Normal%20Extension), and \begin{align*} T(\lambda(a\otimes b))=\tau(\lambda a)\operatorname{Tr}(b)=\lambda\tau(a)\operatorname{Tr}(b)=\lambda T(a\otimes b) \end{align*} for $\lambda\ge 0$. For elementary $x\otimes y\in N$, \begin{align*} T((x\otimes y)^*(x\otimes y))=T(x^*x\otimes y^*y)=\tau(x^*x)\operatorname{Tr}(y^*y). \end{align*} Using the tracial identities for $\tau$ and $\operatorname{Tr}$, \begin{align*} \tau(x^*x)\operatorname{Tr}(y^*y)=\tau(xx^*)\operatorname{Tr}(yy^*)=T(xx^*\otimes yy^*)=T((x\otimes y)(x\otimes y)^*). \end{align*} Normality, faithfulness, and semifiniteness pass to the normal tensor-product trace because $\tau$ is faithful normal finite and $\operatorname{Tr}$ is faithful normal semifinite. The algebra $N$ is a factor: by the standard centre formula for tensor products of von Neumann algebras, \begin{align*} Z(N)=Z(M)\overline{\otimes}Z(B(\ell^2))=\mathbb C1\overline{\otimes}\mathbb C1=\mathbb C(1\otimes 1). \end{align*} It has no nonzero minimal projections because $M$ has none, and tensoring with $B(\ell^2)$ does not create minimal projections. Also the identity is infinite under the trace, since \begin{align*} T(1\otimes 1)=\tau(1)\operatorname{Tr}(1)=1\cdot\infty=\infty. \end{align*} Thus $N$ is a Type $II_\infty$ factor with faithful normal semifinite trace $T=\tau\otimes\operatorname{Tr}$. If $p\in M$ is a projection, then $p\otimes e_{11}$ is a projection because \begin{align*} (p\otimes e_{11})^2=p^2\otimes e_{11}^2=p\otimes e_{11} \end{align*} and \begin{align*} (p\otimes e_{11})^*=p^*\otimes e_{11}^*=p\otimes e_{11}. \end{align*} Its trace is finite and equals the original Type $II_1$ dimension: \begin{align*} T(p\otimes e_{11})=\tau(p)\operatorname{Tr}(e_{11})=\tau(p)\cdot 1=\tau(p). \end{align*} In contrast, \begin{align*} T(1\otimes 1)=\infty. \end{align*} So amplification keeps the finite trace scale of $M$ inside rank-one corners while making the whole identity have infinite trace. [/example] The amplification example contains finite-trace projections such as $1\otimes e_{11}$, and these projections should recover the finite theory when used as identities of corners. This reduction is essential because many semifinite arguments first prove a statement in finite corners and then assemble the result by normality and semifiniteness. The next theorem makes that reduction precise. [quotetheorem:9298] [citeproof:9298] The finite-trace and nonzero assumptions are both essential. If $p=0$, the corner has zero identity and is not a Type $II_1$ factor; if $\operatorname{Tr}(p)=\infty$, the formula for $\tau_p$ cannot normalize the identity of $pNp$ to have trace $1$. Factoriality is also part of the mechanism: corners of non-factors can retain nontrivial centre, so a finite-trace corner in a direct sum of semifinite algebras need not be a factor. The no-minimal-projections hypothesis rules out Type I behaviour inside the corner; for instance, in $B(\ell^2)$ the finite-rank projection onto an $n$-dimensional subspace has corner $M_n(\mathbb C)$, which contains minimal projections and is Type I rather than Type $II_1$. In the amplification $M\overline{\otimes}B(\ell^2)$, the corner cut out by $1\otimes 1$ is the whole Type $II_\infty$ algebra, not a finite factor, while the corner cut out by $1\otimes e_{11}$ recovers the Type $II_1$ algebra $M$. Thus the theorem does not say that every corner of a Type $II_\infty$ factor is finite; it identifies exactly the finite-trace corners in the Type II factorial setting as the places where Type $II_1$ theory applies. This result is the practical link between the finite and semifinite parts of the theory. Many arguments in Type $II_\infty$ algebras are reduced to finite corners, proved with the Type $II_1$ trace, and then reassembled using semifiniteness. [example: Finite Corners of an Amplification] Let $N=M\overline{\otimes}B(\ell^2)$, where $M$ is a Type $II_1$ factor with trace $\tau$, and let $e_{11}$ be the rank-one projection onto $\mathbb C e_1$. Put $p=1\otimes e_{11}$. For $x\in M$, the element $x\otimes e_{11}$ belongs to $pNp$ because \begin{align*} p(x\otimes e_{11})p=(1x1)\otimes(e_{11}e_{11}e_{11})=x\otimes e_{11}. \end{align*} Conversely, every elementary tensor in the corner has the form \begin{align*} p(a\otimes b)p=a\otimes e_{11}be_{11}. \end{align*} Since $e_{11}be_{11}=(be_1,e_1)e_{11}$, this equals \begin{align*} p(a\otimes b)p=(be_1,e_1)a\otimes e_{11}. \end{align*} Thus the corner is exactly $M\otimes \mathbb C e_{11}$, and the map \begin{align*} \Phi:M\to pNp,\qquad \Phi(x)=x\otimes e_{11} \end{align*} is a $*$-isomorphism: for $x,y\in M$, \begin{align*} \Phi(xy)=xy\otimes e_{11}=(x\otimes e_{11})(y\otimes e_{11})=\Phi(x)\Phi(y) \end{align*} and \begin{align*} \Phi(x^*)=x^*\otimes e_{11}=(x\otimes e_{11})^*=\Phi(x)^*. \end{align*} Let $T=\tau\otimes\operatorname{Tr}$ be the semifinite trace on $N$. Since $\operatorname{Tr}(e_{11})=1$, the trace of the corner identity is \begin{align*} T(p)=T(1\otimes e_{11})=\tau(1)\operatorname{Tr}(e_{11})=1. \end{align*} Therefore the normalized corner trace is just $T$ restricted to $pNp$. For $x\in M_+$, \begin{align*} T(\Phi(x))=T(x\otimes e_{11})=\tau(x)\operatorname{Tr}(e_{11})=\tau(x). \end{align*} So the rank-one corner recovers $M$ with exactly its original trace. More generally, let $f_k=e_{11}+\cdots+e_{kk}$. Then $1\otimes f_k$ has trace \begin{align*} T(1\otimes f_k)=\tau(1)\operatorname{Tr}(f_k)=k. \end{align*} The corner $(1\otimes f_k)N(1\otimes f_k)$ identifies with $M\overline{\otimes}M_k(\mathbb C)$, and its normalized trace is \begin{align*} \frac{1}{k}T=\tau\otimes\left(\frac{1}{k}\operatorname{Tr}_{M_k}\right). \end{align*} Thus finite-rank corners of the amplification are finite matrix amplifications of $M$, with the trace normalized by the rank of the projection cutting out the corner. [/example] The chapter therefore completes the passage from equivalence of projections to numerical dimension. Type $II_1$ factors have a unique normalized trace and projection dimensions filling $[0,1]$; Type $II_\infty$ factors replace the state by a semifinite trace and extend the same comparison theory to an infinite scale. Type II algebras complete the passage from projection equivalence to numerical dimension, but the trace machinery also marks its own limit. The next chapter examines what happens when that numerical scale disappears entirely, and modular structure replaces trace as the main invariant. # 9. Type III Algebras and the Absence of Trace After Chapter 8 developed the finite and semifinite trace cases, this chapter completes the type classification by treating the case where trace theory fails. In Types I and II, traces measure projections and make cyclic symmetry $\tau(xy)=\tau(yx)$ available as a structural tool. Type III algebras are the part of the theory where that measuring device disappears: nonzero projections cannot carry finite trace-like mass, and the replacement symmetry is implemented by modular theory rather than by a trace. ## Proper Infiniteness and the Failure of Semifinite Traces What prevents a von Neumann algebra from admitting a trace that sees every nonzero positive element? The obstruction is not just the existence of an infinite projection, since semifinite algebras such as $B(H)$ on an infinite-dimensional Hilbert space contain many infinite projections while still have a faithful normal semifinite trace. The sharper obstruction is that the algebra can be infinite at every nonzero corner. Recall first the comparison language for projections. If $M\subseteq B(H)$ is a von Neumann algebra and $p,q\in M$ are projections, then $p\sim q$ means there is a partial isometry $v\in M$ with $v^*v=p$ and $vv^*=q$, while $p\precsim q$ means $p\sim q_0$ for some projection $q_0\le q$. [definition: Infinite Projection] A projection $p\in M$ is infinite if there exists a projection $q<p$ such that $q\sim p$. [/definition] This definition identifies the projections that can absorb a proper part of themselves, so it gives the comparison-theoretic obstruction to finite trace values. To use it as a classification tool, we also need a name for the projections where this obstruction is absent. This motivates defining finite projections as the projections that cannot be compressed onto a proper subprojection by a partial isometry in $M$. [definition: Finite Projection] A projection $p\in M$ is finite if it is not infinite. [/definition] The point of this terminology is that finiteness is intrinsic to $M$, not to the Hilbert space dimension of $pH$. A projection may have infinite-dimensional range and still be finite in an abelian von Neumann algebra, because abelian algebras have no nontrivial partial isometries moving a set onto a proper subset inside the algebra. [example: Finite Projection With Infinite Hilbert Space Range] Let $M=L^\infty([0,1])$ act on $L^2([0,1])$ by multiplication, and let $p=\mathbb{1}_{[0,1]}$, so $p$ is the identity projection. The range is $pL^2([0,1])=L^2([0,1])$, which is infinite-dimensional because the functions $e_n(t)=e^{2\pi int}$ for $n\in\mathbb Z$ are nonzero and mutually orthogonal. We show that $p$ is finite in $M$. Every projection in $L^\infty([0,1])$ is multiplication by a characteristic function: if $e=e^*=e^2$, then $e(t)\in\{0,1\}$ for almost every $t$, so $e=\mathbb{1}_E$ almost everywhere for some measurable set $E$. If two projections $e,f\in M$ are Murray-von Neumann equivalent, then there is a partial isometry $v\in M$ with $v^*v=e$ and $vv^*=f$. Since $M$ is abelian, $v^*v=vv^*$, hence \begin{align*} e=v^*v=vv^*=f. \end{align*} Thus Murray-von Neumann equivalence of projections inside this abelian algebra is equality. Therefore no proper subprojection $q<p$ can satisfy $q\sim p$, because $q\sim p$ would force $q=p$. Hence $p$ is finite in $M$ even though its Hilbert-space range is infinite-dimensional. [/example] The example shows that finiteness depends on the available partial isometries, not on Hilbert space size alone. For the trace obstruction needed in Type III theory, ordinary infiniteness is still too weak because it can occur in a single corner while other corners remain finite. The next notion rules out finite behaviour after passing to central pieces and captures the idea that a projection contains several disjoint copies of itself. [definition: Properly Infinite Projection] A projection $p\in M$ is properly infinite if there exist mutually orthogonal projections $p_1,p_2\le p$ such that $p_1\sim p$ and $p_2\sim p$. [/definition] Proper infiniteness means that $p$ contains two disjoint internal copies of itself. For factors this is close to being infinite, but for algebras with centre it is the central version that interacts best with traces. [example: The Identity of an Infinite Type I Factor] Let $H$ be a separable infinite-dimensional Hilbert space and let $M=B(H)$. Choose an orthonormal basis $(e_n)_{n\ge 1}$ for $H$, and set \begin{align*} H_1=\overline{\operatorname{span}}\{e_{2n-1}:n\ge 1\}, \qquad H_2=\overline{\operatorname{span}}\{e_{2n}:n\ge 1\}. \end{align*} Then $H=H_1\oplus H_2$, and the orthogonal projections $p_1$ and $p_2$ onto $H_1$ and $H_2$ satisfy $p_1p_2=0$ and $p_1,p_2\le 1$. Define isometries $v_1,v_2:H\to H$ on the basis by \begin{align*} v_1e_n=e_{2n-1}, \qquad v_2e_n=e_{2n}. \end{align*} For $i=1,2$, the map $v_i$ preserves inner products on the orthonormal basis and hence extends to an isometry on $H$, so $v_i^*v_i=1$. The range of $v_1$ is $H_1$, and the range of $v_2$ is $H_2$, so \begin{align*} v_1v_1^*=p_1, \qquad v_2v_2^*=p_2. \end{align*} Thus $p_1\sim 1$ and $p_2\sim 1$ inside $B(H)$. Since $p_1$ and $p_2$ are orthogonal subprojections of $1$, the identity projection contains two disjoint Murray-von Neumann equivalent copies of itself. Therefore $1$ is properly infinite in $B(H)$. [/example] This example also shows why proper infiniteness alone cannot characterize Type III: $B(H)$ is properly infinite but has the usual faithful normal semifinite trace on positive operators, \begin{align*} \tau(T)=\sum_n (Te_n,e_n)_H, \end{align*} independent of the chosen orthonormal basis when interpreted in $[0,\infty]$. We therefore need the exact trace notion whose failure will distinguish Type III algebras from the semifinite properly infinite cases. The key extra condition is semifiniteness, which says that finite trace pieces occur densely enough to detect nonzero positive operators. [definition: Faithful Normal Semifinite Trace] A faithful normal semifinite trace on a von Neumann algebra $M$ is a map $\tau:M_+\to[0,\infty]$ such that: 1. $\tau(x+y)=\tau(x)+\tau(y)$ and $\tau(\lambda x)=\lambda\tau(x)$ for $x,y\in M_+$ and $\lambda\ge 0$; 2. $\tau(a^*a)=\tau(aa^*)$ for every $a\in M$; 3. $\tau(x)=0$ implies $x=0$ for $x\in M_+$; 4. $\tau(\sup_i x_i)=\sup_i\tau(x_i)$ for every increasing bounded net $(x_i)$ in $M_+$; 5. for every nonzero $x\in M_+$ there exists $0\le y\le x$ with $y\ne0$ and $\tau(y)<\infty$. [/definition] The trace identity forces Murray-von Neumann equivalent projections to have the same trace value. That observation turns the comparison relation on projections into a numerical obstruction. The next theorem records the basic incompatibility between infiniteness and finite trace, which will be used to rule out semifinite traces when no finite projection remains. [quotetheorem:9299] [citeproof:9299] The theorem does not say that semifinite traces forbid infinite projections; it says that finite trace lives only on finite pieces. The infiniteness hypothesis is essential: a rank-one projection in $B(H)$ is finite and has finite trace. Faithfulness is also essential, since the zero trace would assign finite value to every projection while detecting nothing. Semifiniteness is not used in this short argument itself, but it is part of the ambient trace notion needed below, where finite-trace subpieces are extracted from positive operators. A semifinite trace can still exist when the algebra has enough finite projections below its positive operators. If every nonzero projection is already infinite, the spectral theorem converts semifiniteness into a direct contradiction. [quotetheorem:9300] [citeproof:9300] This is the first precise form of the slogan that Type III algebras have no trace. The hypothesis is much stronger than proper infiniteness of the identity: $B(H)$ has properly infinite identity, but its rank-one projections are finite and support the usual semifinite trace. The proof also gives a practical test for trace arguments: use semifiniteness to find a positive element of finite trace, then pass to a nonzero spectral projection $1_{[\varepsilon,\infty)}(x)$; if comparison theory says every such projection is infinite, the trace argument cannot apply. The statement is stronger than the absence of a finite trace state: the algebra has no faithful normal semifinite trace, so the standard noncommutative integration theory based on traces has no direct foothold. ## Type III Factors and the Absence of Finite Projections What distinguishes a Type III factor from the properly infinite factors already seen in Type I$_\infty$ and Type II$_\infty$? In the semifinite cases, although the identity is infinite, there are still nonzero finite projections: rank-one projections in $B(H)$ and finite-trace projections in a Type II$_\infty$ factor. Type III factors are exactly the factors where this finite part has vanished. [definition: Type III Factor] A factor $M$ is of Type III if $M$ has no nonzero finite projections. [/definition] Because a factor has centre $\mathbb C1$, this definition is equivalent to saying that every nonzero projection is infinite. The formulation in terms of finite projections is often more useful because it contrasts directly with the Type I and Type II cases. [quotetheorem:9301] [citeproof:9301] The cited criterion uses only projection comparison and corners, avoiding any ambiguous trace ideal. The connection with traces comes from the previous chapter: finite factors carry faithful normal tracial states, while the theorem above says that a Type III factor has no nonzero corner on which such finite trace theory can begin. This also explains the role of the factor hypothesis. For a concrete nonfactor, take \begin{align*} M=M_2(\mathbb C)\oplus B(\ell^2). \end{align*} The central projection $z_1=(1,0)$ supports a finite algebra, while $z_2=(0,1)$ supports a properly infinite semifinite algebra. Thus finite projections and infinite projections coexist in one von Neumann algebra because they lie on different central summands. The theorem is not saying that a properly infinite identity forces Type III behaviour, nor is it a classification of arbitrary nonfactors without central decomposition. It says that, once the centre has collapsed to $\mathbb C1$, the absence of a single nonzero finite projection is equivalent to the absence of every finite corner. [example: Contrast With Semifinite Factors] Let $H$ be separable and infinite-dimensional, and let $M=B(H)$ with the usual semifinite trace $\tau(T)=\sum_n(Te_n,e_n)_H$. If $e$ is the rank-one projection onto $\mathbb C e_1$, then \begin{align*} \tau(e)=(ee_1,e_1)_H+\sum_{n\ge 2}(ee_n,e_n)_H=1+\sum_{n\ge 2}0=1. \end{align*} The projection $e$ is finite: if $q\le e$, then $qH\subseteq eH=\mathbb C e_1$, so $q=0$ or $q=e$; and $q\sim e$ cannot hold when $q=0$, since $v^*v=e$ and $vv^*=0$ would force $v=0$ and hence $e=0$. Thus no proper subprojection of $e$ is equivalent to $e$. The identity in $B(H)$ is infinite. With respect to the basis $(e_n)_{n\ge1}$, define $v e_n=e_{n+1}$. Then $v$ is an isometry, so $v^*v=1$, while its range is $(\mathbb C e_1)^\perp$, so \begin{align*} vv^*=1-e. \end{align*} Since $1-e<1$ and $1-e\sim 1$, the identity is infinite. This shows that a semifinite factor can contain infinite projections while still having finite projections of finite trace. In a Type II$_\infty$ factor, the faithful normal semifinite trace has $\tau(1)=\infty$, but the continuous dimension range gives projections $p_s$ with $\tau(p_s)=s$ for every $s\in[0,\infty)$. Thus the identity is infinite, while finite-trace projections remain available below it. In a Type III factor, every nonzero projection $p$ is infinite by the definition of Type III factor, so *Infinite Projections Have Infinite Faithful Trace* forces $\tau(p)=\infty$ for any faithful normal semifinite trace $\tau$ that could exist. Hence the Type III obstruction is not that the identity is large, but that every nonzero projection is already too infinite to carry finite trace. [/example] This absence also changes how states should be interpreted. In a finite factor, a faithful normal tracial state is compatible with the algebraic symmetry $ab\leftrightarrow ba$. In a Type III factor, normal states exist in abundance, but they are not traces and their failure to be tracial contains structural information. [remark: Normal States Are Not Traces] A Type III factor may have many faithful normal states $\varphi:M\to\mathbb C$. None of them is tracial, because a faithful normal tracial state would be a faithful normal finite trace, and every projection would then have finite trace at most $\varphi(1)=1$. This would contradict the fact that every nonzero projection is infinite. [/remark] The most important examples of Type III algebras arise when localization and infinitely many degrees of freedom are both present. The finite-dimensional intuition from matrices and density operators no longer describes the projection structure. [example: Local Algebras in Relativistic Quantum Field Theory] In algebraic quantum field theory, a bounded spacetime region $O$ is assigned a von Neumann algebra $\mathcal A(O)$ of observables localized in $O$. In the physically standard local models, $\mathcal A(O)$ is a Type III factor, often Type III$_1$, so by *Type III Factor Characterization* every nonzero projection in $\mathcal A(O)$ is infinite. This immediately rules out minimal projections. Indeed, if $e\in \mathcal A(O)$ were a nonzero minimal projection, then every projection $q\le e$ would satisfy $q=0$ or $q=e$. Hence there is no proper subprojection $q<e$ with $q\sim e$, so $e$ would be finite. This contradicts the Type III condition that every nonzero projection is infinite. It also rules out the finite-dimensional density-matrix picture. In a tensor-factor model one would describe a normal state by a density operator $\rho$ and a trace formula $\varphi(x)=\operatorname{Tr}(\rho x)$. But a Type III factor has no nonzero finite projections, and the trace obstruction in *No Semifinite Trace When All Nonzero Projections Are Infinite* says that no faithful normal semifinite trace can exist in this situation. Thus the vacuum restricted to $\mathcal A(O)$ is not represented by an ordinary density matrix on a local tensor factor. The algebraic shadow of short-scale entanglement is precisely this projection structure: every nonzero local projection is already infinite, so there are no atoms and no finite trace pieces inside the local algebra. [/example] The canonical mathematical examples were constructed to show that Type III factors are not only a pathology of quantum field theory. They also arise from concrete operator-algebraic constructions. [example: Powers Factors] Fix $0<\lambda<1$ and let $\rho_\lambda\in M_2(\mathbb C)$ be the diagonal density matrix with diagonal entries $(1+\lambda)^{-1}$ and $\lambda(1+\lambda)^{-1}$. Define the one-site state by $\varphi_\lambda(x)=\operatorname{Tr}(\rho_\lambda x)$. If $e_{ij}$ are the standard matrix units, then $e_{12}e_{21}=e_{11}$, so \begin{align*} \varphi_\lambda(e_{12}e_{21})=\varphi_\lambda(e_{11})=(1+\lambda)^{-1}. \end{align*} Also $e_{21}e_{12}=e_{22}$, hence \begin{align*} \varphi_\lambda(e_{21}e_{12})=\varphi_\lambda(e_{22})=\lambda(1+\lambda)^{-1}. \end{align*} Since $0<\lambda<1$, these two values are unequal, so $\varphi_\lambda$ is not tracial. For a simple tensor $x_1\otimes\cdots\otimes x_n$ in the first $n$ tensor factors, the infinite product state is specified by \begin{align*} \Phi_\lambda(x_1\otimes\cdots\otimes x_n)=\varphi_\lambda(x_1)\cdots\varphi_\lambda(x_n). \end{align*} The Powers factor is the von Neumann algebra generated in the GNS representation of this infinite tensor product state. By *Powers' construction theorem*, this von Neumann algebra is a factor of type III$_\lambda$, so it has no nonzero finite projections by the definition of a Type III factor. The parameter $\lambda$ is already visible in the one-site modular action. Since $\rho_\lambda$ has eigenvalue $(1+\lambda)^{-1}$ on the range of $e_{11}$ and eigenvalue $\lambda(1+\lambda)^{-1}$ on the range of $e_{22}$, functional calculus gives \begin{align*} \rho_\lambda^{it}e_{12}\rho_\lambda^{-it}=(1+\lambda)^{-it}\lambda^{-it}(1+\lambda)^{it}e_{12}. \end{align*} Canceling the factors $(1+\lambda)^{-it}(1+\lambda)^{it}=1$ gives \begin{align*} \rho_\lambda^{it}e_{12}\rho_\lambda^{-it}=\lambda^{-it}e_{12}. \end{align*} Similarly, \begin{align*} \rho_\lambda^{it}e_{21}\rho_\lambda^{-it}=\lambda^{it}e_{21}. \end{align*} Thus the modular dynamics records the ratio of the two eigenvalues of the reference state, and that ratio is exactly $\lambda$. The Powers factors therefore give a concrete family where the absence of finite projections is paired with modular data that distinguishes the Type III$_\lambda$ subclasses. [/example] These examples motivate the final section. If trace symmetry disappears, there must be another canonical way to encode how a faithful normal state fails to commute with multiplication. Tomita-Takesaki theory supplies that replacement. ## Modular Automorphisms as the Replacement for Trace Symmetry How can a nontracial state organize a von Neumann algebra? For a trace $\tau$, the equality $\tau(ab)=\tau(ba)$ means that left and right multiplication are balanced. For a faithful normal state $\varphi$ on a Type III algebra, the imbalance is not noise; it produces a canonical one-parameter automorphism group. The cleanest statement uses the standard form coming from the GNS representation. Let $M$ be a von Neumann algebra and let $\varphi$ be a faithful normal state. In the GNS Hilbert space $H_\varphi$, the cyclic vector $\Omega_\varphi$ is also separating for $M$, so the formula below defines a densely defined antilinear operator on the dense subspace $M\Omega_\varphi\subset H_\varphi$. [definition: Modular Operator and Modular Conjugation] Let $M$ be a von Neumann algebra with faithful normal state $\varphi$, represented on its GNS Hilbert space $H_\varphi$ with cyclic separating vector $\Omega_\varphi$. Define the densely defined antilinear operator \begin{align*} S_0:M\Omega_\varphi\subset H_\varphi&\longrightarrow H_\varphi, & S_0(a\Omega_\varphi)&=a^*\Omega_\varphi. \end{align*} Let \begin{align*} S:\operatorname{Dom}(S)\subset H_\varphi&\longrightarrow H_\varphi \end{align*} be the closure of $S_0$. The polar decomposition of $S$ is \begin{align*} S=J_\varphi \Delta_\varphi^{1/2}, \end{align*} where \begin{align*} J_\varphi:H_\varphi&\longrightarrow H_\varphi, & \Delta_\varphi:\operatorname{Dom}(\Delta_\varphi)\subset H_\varphi&\longrightarrow H_\varphi \end{align*} with $J_\varphi$ antiunitary and $\Delta_\varphi$ positive self-adjoint. [/definition] The definition packages the GNS asymmetry of a faithful normal state into two operators, $J_\varphi$ and $\Delta_\varphi$. This setup motivates the next theorem because the construction would not help with Type III algebras unless these Hilbert-space operators acted canonically on $M$ itself. Tomita-Takesaki is needed precisely to prove that the imaginary powers of $\Delta_\varphi$ preserve $M$ and that $J_\varphi$ identifies the commutant. [quotetheorem:9302] The full proof belongs to modular theory and is usually quoted rather than proved in this course. It rests on the [analytic continuation](/page/Analytic%20Continuation) properties of the Tomita algebra, the polar decomposition of the closed antilinear operator $S$, and the relation between $S$, its adjoint, and the commutant action. The cyclic and separating hypotheses are not cosmetic: cyclicity makes $M\Omega$ large enough to define $S$ densely, while separatingness makes the rule $a\Omega\mapsto a^*\Omega$ well-defined. The standard-representation language packages these hypotheses so that the commutant relation $JMJ=M'$ is meaningful. The theorem is the main structural replacement for trace-based symmetry in the Type III setting. [definition: Modular Automorphism Group] Let $M$ be a von Neumann algebra and let $\varphi$ be a faithful normal state. The modular automorphism group of $\varphi$ is the one-parameter group $(\sigma_t^\varphi)_{t\in\mathbb R}$ of maps \begin{align*} \sigma_t^\varphi:M&\longrightarrow M, & \sigma_t^\varphi(x)&=\Delta_\varphi^{it}x\Delta_\varphi^{-it}. \end{align*} [/definition] This group is canonical once the state is chosen. The dependence on $\varphi$ is essential: Type III theory studies not only the existence of these flows, but also what remains invariant when the faithful normal state changes. [quotetheorem:9303] [citeproof:9303] The invariance theorem shows that modular automorphisms preserve the chosen state, so they play part of the role that trace-preserving symmetries played in finite theory. However, preservation of a state is weaker than cyclic trace symmetry, and the hypotheses are doing real work. If $\varphi$ is not faithful, the GNS vector need not be separating: for the vector state $\varphi(x)=(xe_1,e_1)_{\mathbb C^2}$ on $M_2(\mathbb C)$, the matrix unit $e_{12}$ satisfies $e_{12}e_1=0$ while $e_{12}^*e_1=e_2$, so the rule $a\Omega\mapsto a^*\Omega$ is not well-defined on the quotient determined by the vector. If normality is dropped, a singular state on $B(\ell^2)$ is not represented by a density operator on the original Hilbert space, and the resulting GNS representation is not the normal representation used in von Neumann algebra modular theory. Outside this faithful normal modular setup, a state-preserving flow is only an extra dynamical choice; it need not be the canonical modular flow. To see exactly where the finite theory sits inside modular theory, we next examine what happens when the state is tracial. [quotetheorem:9304] [citeproof:9304] This theorem gives the conceptual bridge from finite theory to Type III theory. The tracial hypothesis cannot be weakened to faithfulness alone. In $M_n(\mathbb C)$, let $\rho$ be an invertible density matrix that is not a scalar multiple of the identity and define $\varphi(x)=\operatorname{Tr}(\rho x)$. Then $\varphi$ is faithful and normal but not tracial, and its modular group is \begin{align*} \sigma_t^\varphi(x)=\rho^{it}x\rho^{-it}. \end{align*} This flow is nonidentity whenever $x$ does not commute with $\rho$. Thus even a finite-dimensional algebra can have nonzero modular dynamics if the state is nontracial. In finite algebras the modular flow collapses precisely for faithful normal tracial states, while in Type III algebras the flow is generally nonzero and becomes part of the invariant structure. [example: Finite Systems Versus Type III Systems] For the normalized trace on $M_n(\mathbb C)$, \begin{align*} \tau(ab)=n^{-1}\operatorname{Tr}(ab)=n^{-1}\operatorname{Tr}(ba)=\tau(ba). \end{align*} Thus $\tau$ is tracial, so by *Tracial States Have Identity Modular Group* its modular automorphism group satisfies $\sigma_t^\tau(x)=x$ for every $x\in M_n(\mathbb C)$ and $t\in\mathbb R$. Now let $\rho$ be an invertible density matrix and define $\varphi(x)=\operatorname{Tr}(\rho x)$. The finite-dimensional modular formula is \begin{align*} \sigma_t^\varphi(x)=\rho^{it}x\rho^{-it}. \end{align*} Since $\rho$ is positive and invertible, functional calculus gives $\rho^{it}$ unitary, because \begin{align*} (\rho^{it})^*\rho^{it}=\rho^{-it}\rho^{it}=1. \end{align*} Hence $x\mapsto \rho^{it}x\rho^{-it}$ is an inner $*$-automorphism. If $\rho$ is diagonal with eigenvalues $\alpha_1,\ldots,\alpha_n>0$, then for the matrix unit $e_{ij}$, \begin{align*} \rho^{it}e_{ij}\rho^{-it}=\alpha_i^{it}\alpha_j^{-it}e_{ij}=(\alpha_i/\alpha_j)^{it}e_{ij}. \end{align*} So whenever $\alpha_i\ne\alpha_j$, the modular flow moves $e_{ij}$ for values of $t$ with $(\alpha_i/\alpha_j)^{it}\ne1$. In Type III factors this modular dynamics cannot be eliminated by replacing the state with a trace: by *No Semifinite Trace When All Nonzero Projections Are Infinite*, no faithful normal semifinite trace exists there. [/example] The absence of trace therefore does not leave Type III algebras structureless. It shifts the centre of the theory from numerical dimension functions to state-dependent dynamics. Later refinements of Type III classification, such as the Connes spectrum and the subclasses Type III$_\lambda$, extract invariants from modular automorphism groups rather than from traces. Type III algebras show that the familiar trace-based notion of dimension is not universal, so the theory must be read through states and modular dynamics instead. The final part of the course brings the operator-algebraic structure back to quantum mechanics, where these same notions describe observables, measurements, and physical symmetry. # 10. States, Observables, and Quantum Mechanics This chapter translates the operator-algebraic structure developed in Chapters 1--9 into the language of quantum mechanics. The guiding point is that a von Neumann algebra $M \subseteq \mathcal{L}(H)$ is not only a closed algebra of operators: it is a proposed algebra of observables for a physical system represented on a Hilbert space $H$. The weak and ultraweak closures are essential here, because spectral projections, normal states, and compatibility of measurements are all expressed by limits and commutants rather than by norm approximation alone. ## Observables and Affiliation Which operators should count as observables when quantum measurements can have unbounded numerical outcomes? Bounded self-adjoint operators in $M$ are the simplest observables, but position, momentum, and Hamiltonians are typically unbounded. The von Neumann algebraic answer is to retain control of such operators through their spectral projections. A bounded measurement with finitely many possible outcomes is modeled by a self-adjoint element of the algebra. [definition: Bounded Observable] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra on a Hilbert space $H$. A bounded observable in $M$ is an element $a \in M$ such that $a=a^*$. [/definition] The spectral theorem assigns to such an $a$ a projection-valued measure $E_a$ on $\mathbb R$, and the condition $a \in M$ forces every spectral projection $E_a(B)$ to lie in $M$. The following finite-dimensional example motivates the later passage from operators to their spectral projections. [example: Finite Quantum System Observables] For $H=\mathbb C^n$ and $M=M_n(\mathbb C)$, a bounded observable is exactly a matrix $a\in M_n(\mathbb C)$ with $a=a^*$, so it is Hermitian. Let the distinct eigenvalues of $a$ be $\lambda_1,\dots,\lambda_r$, and let $V_j=\ker(a-\lambda_j I)$ be the corresponding eigenspace. The finite-dimensional spectral theorem gives an orthogonal direct sum \begin{align*} \mathbb C^n=V_1\oplus\cdots\oplus V_r. \end{align*} Let $p_j$ be the orthogonal projection onto $V_j$. Then $p_j=p_j^*$, $p_j^2=p_j$, and $p_jp_k=0$ when $j\ne k$, while \begin{align*} p_1+\cdots+p_r=I. \end{align*} For any vector $\xi\in\mathbb C^n$, write \begin{align*} \xi=p_1\xi+\cdots+p_r\xi. \end{align*} Since $p_j\xi\in V_j$, the definition of $V_j$ gives \begin{align*} a(p_j\xi)=\lambda_jp_j\xi. \end{align*} Using linearity of $a$, we therefore get \begin{align*} a\xi=a(p_1\xi+\cdots+p_r\xi). \end{align*} By linearity again, \begin{align*} a\xi=a(p_1\xi)+\cdots+a(p_r\xi). \end{align*} Substituting the eigenvalue equation on each eigenspace gives \begin{align*} a\xi=\lambda_1p_1\xi+\cdots+\lambda_rp_r\xi. \end{align*} Thus \begin{align*} a\xi=\left(\sum_{j=1}^r\lambda_jp_j\right)\xi. \end{align*} Because this holds for every $\xi\in\mathbb C^n$, \begin{align*} a=\sum_{j=1}^r\lambda_jp_j. \end{align*} Each $p_j$ is an $n\times n$ matrix, hence $p_j\in M_n(\mathbb C)$. Conversely, the displayed formula reconstructs $a$ from the eigenvalues and the projections $p_j$, so in finite dimensions the observable and its measurement projections determine the same algebraic data. [/example] The finite-dimensional example contributes the principle that measurement events are projections inside the algebra. In infinite-dimensional systems, unbounded observables cannot be elements of $M$, since $M$ consists of bounded operators, so this principle motivates a definition that controls an unbounded operator through its compatibility with $M'$. [definition: Affiliated Self-Adjoint Operator] Let $M \subseteq \mathcal{L}(H)$ be a von Neumann algebra. A densely defined self-adjoint operator \begin{align*} T:\operatorname{dom}(T)\subset H\to H \end{align*} is affiliated with $M$, said to be affiliated with $M$, if every unitary $u \in M'$ satisfies \begin{align*} uT \subseteq Tu. \end{align*} [/definition] The relation $uT\subseteq Tu$ means that $u$ maps the domain of $T$ into itself and $Tu\xi=uT\xi$ for all $\xi\in\operatorname{dom}(T)$. For an unbounded operator this domain condition is awkward to verify directly, and it is not expressed in the projection language of von Neumann algebras. To use affiliated operators in examples, one needs a criterion stated only in terms of bounded projections. Since a self-adjoint operator is measured through its spectral projections, the useful test is whether those spectral projection events belong to $M$. [quotetheorem:9305] [citeproof:9305] This criterion is the practical test for whether an unbounded observable belongs to the measurable structure encoded by $M$. The self-adjointness hypothesis is essential because it gives a projection-valued spectral measure on $\mathbb R$; for instance, the symmetric operator $-i\,d/dx$ on $C_c^\infty(0,\infty)\subset L^2(0,\infty)$ is not self-adjoint and therefore has no spectral projection-valued measure on $\mathbb R$ to test against $M$. The result does not say that $T$ is bounded or that $T\in M$, since an affiliated operator may be unbounded and therefore cannot be an element of the bounded algebra $M$. For example, multiplication by $x$ on $L^2(\mathbb R)$ is affiliated with $L^\infty(\mathbb R)$ but is not a bounded multiplication operator. The next example applies the criterion in the commutative model, where affiliated operators become ordinary measurable functions. [example: Multiplication Observables] Let $(X,\mathcal E,\mu)$ be a measure space, let $M=L^\infty(X,\mu)$ act on $H=L^2(X,\mu)$ by multiplication, and let $f:X\to\mathbb R$ be measurable and finite almost everywhere. Define \begin{align*} \operatorname{dom}(T_f)=\{\xi\in L^2(X,\mu):f\xi\in L^2(X,\mu)\} \end{align*} and $T_f\xi=f\xi$ on this domain. Since $f$ is real-valued almost everywhere, for $\xi,\eta\in\operatorname{dom}(T_f)$ we have \begin{align*} (T_f\xi,\eta)=\int_X f(x)\xi(x)\overline{\eta(x)}\,d\mu(x) \end{align*} and also \begin{align*} (\xi,T_f\eta)=\int_X \xi(x)\overline{f(x)\eta(x)}\,d\mu(x)=\int_X \xi(x)f(x)\overline{\eta(x)}\,d\mu(x). \end{align*} The two integrals are equal because $f(x)=\overline{f(x)}$ almost everywhere, so $T_f$ is symmetric. The standard adjoint computation for maximal multiplication operators gives $\operatorname{dom}(T_f^*)=\{\eta\in L^2(X,\mu):f\eta\in L^2(X,\mu)\}$ and $T_f^*\eta=f\eta$, hence $T_f^*=T_f$. For a Borel set $B\subseteq\mathbb R$, define $P_B$ on $L^2(X,\mu)$ by \begin{align*} (P_B\xi)(x)=\mathbb{1}_{f^{-1}(B)}(x)\xi(x). \end{align*} Since $\mathbb{1}_{f^{-1}(B)}$ is bounded by $1$, this is a bounded multiplication operator. It is a projection because \begin{align*} P_B^2\xi=\mathbb{1}_{f^{-1}(B)}^2\xi=\mathbb{1}_{f^{-1}(B)}\xi=P_B\xi, \end{align*} and it is self-adjoint because $\mathbb{1}_{f^{-1}(B)}$ is real-valued. Thus $P_B$ is multiplication by an element of $L^\infty(X,\mu)$, so $P_B\in M$. The spectral measure of the real multiplication operator $T_f$ is exactly $E_{T_f}(B)=P_B$, because applying a bounded Borel function $g$ to $T_f$ gives multiplication by $g\circ f$; for $g=\mathbb{1}_B$, this is multiplication by $\mathbb{1}_B\circ f=\mathbb{1}_{f^{-1}(B)}$. Therefore every spectral projection of $T_f$ belongs to $M$, and by *[Spectral Projection Criterion for Affiliation](/theorems/9305)*, $T_f$ is affiliated with $M$. In this commutative model, an unbounded observable is therefore just a real measurable function, and its measurement events are the measurable subsets $f^{-1}(B)$ encoded as projections in $L^\infty(X,\mu)$. [/example] ## Normal States and Expectation Values How does a state assign probabilities and expected measurement outcomes without losing the ultraweak structure of the algebra? In a von Neumann algebra, the physically stable states are the normal states: they are precisely the states continuous for the topology in which increasing nets of projections have their natural limits. The probability assigned to a projection should be compatible with countable decompositions of mutually exclusive events. This motivates restricting attention from arbitrary algebraic states to ultraweakly continuous states. [definition: Normal State] Let $M$ be a von Neumann algebra. A normal state on $M$ is a linear functional $\phi:M\to\mathbb C$ such that $\phi(1)=1$, $\phi(a^*a)\ge 0$ for all $a\in M$, and $\phi$ is ultraweakly continuous. [/definition] Normality ensures stability under the limits already built into a von Neumann algebra, but the definition is topological and does not immediately look like a probability axiom. For measurement events, the basic test is additivity over mutually exclusive projections: the probability of a disjoint union should be the sum of the probabilities. In a general von Neumann algebra the relevant disjoint unions may be arbitrary orthogonal joins, so the probabilistic content of normality is complete additivity on projection events. [quotetheorem:9306] [citeproof:9306] Complete additivity is the correct projection-level analogue of countable additivity for arbitrary von Neumann algebras. The passage from sequences to arbitrary orthogonal families matters because a general von Neumann algebra need not be countably decomposable; in such an algebra, a state can behave well on every countable orthogonal family but still fail to be normal on larger joins. In countably decomposable settings, such as many separably represented algebras used in elementary quantum mechanics, the sequential version often suffices, but the intrinsic theorem is formulated with arbitrary families. This turns spectral projections into probability measures compatible with the ultraweak structure. Since a bounded observable is recovered from its spectral projections by integration, this motivates defining expectation values by evaluating the state on the observable itself. [definition: Expectation Value] Let $M$ be a von Neumann algebra, let $\phi$ be a normal state on $M$, and let $a=a^*\in M$. The expectation value of $a$ in the state $\phi$ is the number $\phi(a)\in\mathbb R$. [/definition] For an unbounded observable $T$ affiliated with $M$, the same idea applies when the first moment exists. The probability law is $\mu_T^\phi(B)=\phi(E_T(B))$, and the expectation is \begin{align*} \int_{\mathbb R}t\,d\mu_T^\phi(t) \end{align*} whenever this integral is finite. The finite-dimensional case motivates the density-matrix notation familiar from quantum mechanics. [example: Density Matrices on a Finite System] Let $M=M_n(\mathbb C)$. A density matrix is a matrix $\rho\in M_n(\mathbb C)$ with $\rho\ge 0$ and $\operatorname{Tr}(\rho)=1$, and it defines a state by \begin{align*} \phi_\rho(a)=\operatorname{Tr}(\rho a). \end{align*} Indeed, $\phi_\rho(I)=\operatorname{Tr}(\rho I)=\operatorname{Tr}(\rho)=1$, and for $a\in M_n(\mathbb C)$, \begin{align*} \phi_\rho(a^*a)=\operatorname{Tr}(\rho a^*a). \end{align*} Since $\rho\ge 0$, let $\rho^{1/2}$ be its positive square root. Using cyclicity of the finite-dimensional trace, \begin{align*} \operatorname{Tr}(\rho a^*a)=\operatorname{Tr}(\rho^{1/2}a^*a\rho^{1/2}). \end{align*} Again by cyclicity, \begin{align*} \operatorname{Tr}(\rho^{1/2}a^*a\rho^{1/2})=\operatorname{Tr}((a\rho^{1/2})^*(a\rho^{1/2})). \end{align*} The last trace is a sum of squared absolute values of matrix entries, so it is nonnegative. Now let $a=a^*$ have spectral decomposition \begin{align*} a=\sum_j\lambda_jp_j. \end{align*} Substituting this expression into the trace formula gives \begin{align*} \phi_\rho(a)=\operatorname{Tr}\left(\rho\sum_j\lambda_jp_j\right). \end{align*} Matrix multiplication distributes over finite sums, so \begin{align*} \rho\sum_j\lambda_jp_j=\sum_j\lambda_j\rho p_j. \end{align*} [Linearity of the trace](/theorems/7814) then gives \begin{align*} \operatorname{Tr}\left(\sum_j\lambda_j\rho p_j\right)=\sum_j\lambda_j\operatorname{Tr}(\rho p_j). \end{align*} Therefore \begin{align*} \phi_\rho(a)=\sum_j\lambda_j\operatorname{Tr}(\rho p_j). \end{align*} For each spectral projection $p_j$, the number $\operatorname{Tr}(\rho p_j)$ is nonnegative because \begin{align*} \operatorname{Tr}(\rho p_j)=\operatorname{Tr}(p_j\rho p_j)=\operatorname{Tr}((p_j\rho^{1/2})^*(p_j\rho^{1/2})). \end{align*} Also, \begin{align*} \sum_j\operatorname{Tr}(\rho p_j)=\operatorname{Tr}\left(\rho\sum_jp_j\right). \end{align*} Since the spectral projections sum to the identity, \begin{align*} \operatorname{Tr}\left(\rho\sum_jp_j\right)=\operatorname{Tr}(\rho I)=1. \end{align*} Thus the coefficients $\operatorname{Tr}(\rho p_j)$ form a probability distribution on the possible measurement outcomes $\lambda_j$, and the expectation is the weighted sum of those outcomes. [/example] Density matrices explain all finite-dimensional normal states, but infinite-dimensional Hilbert spaces introduce a convergence problem: a positive operator with ordinary matrix trace need not exist, and arbitrary states on $\mathcal{L}(H)$ need not be given by summing diagonal matrix coefficients. The right replacement for a finite density matrix is a positive trace-class operator, because trace-class operators are exactly those for which the trace pairing with every bounded operator is well-defined and normal. Thus the question is precisely which states on $\mathcal{L}(H)$ arise from such trace-class density operators. [quotetheorem:9307] [citeproof:9307] The representation theorem shows that the familiar density matrix formalism is a special case of normal states on the full algebra of bounded operators. The hypothesis that the algebra is $\mathcal{L}(H)$ is doing real work: arbitrary von Neumann algebras have preduals, but trace-class representatives on the original Hilbert space are not part of the intrinsic data of the algebra. For example, if $M=\mathbb C1_H$ on an infinite-dimensional Hilbert space $H$, then $M$ has the single normal state $\phi(\lambda 1_H)=\lambda$, while every density operator $\rho$ on $H$ with $\operatorname{Tr}(\rho)=1$ gives the same restricted formula $\phi(\lambda 1_H)=\operatorname{Tr}(\rho\,\lambda 1_H)$. Thus the density operator is highly non-unique and belongs to the chosen ambient representation, not to $M$ itself. The theorem also says nothing about singular states on $\mathcal{L}(H)$, which are states not given by density operators and not continuous for the ultraweak topology. To study a normal state on an arbitrary von Neumann algebra, one needs a representation in which the state becomes a vector state; this motivates the von Neumann version of the GNS construction. [quotetheorem:9308] [citeproof:9308] This form of GNS converts an abstract normal state into a concrete vector state while preserving the von Neumann algebraic continuity needed for spectral projections and commutants. Normality is needed here to ensure that the representation is normal; the purely C*-algebraic GNS construction exists for every state, but it need not preserve ultraweak limits of increasing projection nets. Cyclicity records that the whole representation is generated from the state vector by applying observables, while faithfulness is not automatic: a non-faithful state can annihilate a nonzero projection and the resulting representation then has a nontrivial kernel. With states and expectation values in place, the next question is which observables can be measured together. ## Compatible Observables and Superselection When can two collections of observables be measured together? The operator-algebraic condition is commutation: an observable compatible with all observables in $M$ belongs to the commutant $M'$. This section interprets the commutant and double commutant as measurement-theoretic objects. A compatible measurement should not disturb the projection events of the algebra under discussion. This motivates identifying the algebra of all bounded observables compatible with $M$ as its commutant. [definition: Compatible Observable Algebra] Let $M\subseteq\mathcal{L}(H)$ be a von Neumann algebra. The compatible observable algebra of $M$ is its commutant \begin{align*} M'=\{b\in\mathcal{L}(H):ba=ab\text{ for all }a\in M\}. \end{align*} [/definition] The definition says that compatible bounded observables commute algebraically with $M$, but measurement compatibility is stated in terms of events, not just products of operators. For self-adjoint observables the events are spectral projections, and algebraic commutation would not justify the interpretation unless it forced those projection-valued events to commute as well. The needed bridge is a spectral-theoretic result: bounded commuting self-adjoint operators have commuting Borel functional calculi. The point of invoking this result here is to turn the algebraic definition of $M'$ into a statement about measurement events. We need to know that once an observable commutes with the algebra generated by another observable, all Borel yes-or-no questions obtained from their spectral measures are compatible as projections. [quotetheorem:9309] [citeproof:9309] The theorem gives the operational meaning of the commutant: it is the algebra of bounded observables whose spectral events are compatible with those of $M$. Boundedness keeps the ordinary [Borel functional calculus](/theorems/2696) inside $\mathcal{L}(H)$, and self-adjointness is what supplies projection-valued spectral measures representing measurement events. The statement does not claim that noncommuting observables admit simultaneous sharp measurements; for instance, the usual position and momentum operators have spectral measures that do not commute. For unbounded observables, the corresponding assertion requires the stronger language of commuting spectral measures rather than merely multiplying the unbounded operators on a common domain. The following example motivates the generated measurement algebra associated to a single observable. [example: Measurement Algebra Generated by One Observable] Let $a=a^*\in\mathcal{L}(H)$ and let $M=\{a\}''$. By the bicommutant theorem, $M$ is the von Neumann algebra generated by $a$, so it is the ultraweakly closed unital $*$-algebra obtained from the functional calculus of $a$. If $E_a$ is the spectral measure of $a$, then for every Borel set $B\subseteq\mathbb R$ the bounded Borel function $\mathbb{1}_B$ gives \begin{align*} \mathbb{1}_B(a)=E_a(B). \end{align*} Thus the projection corresponding to the measurement event “the value of $a$ lies in $B$” belongs to $M$. Now let $b=b^*\in\mathcal{L}(H)$ be a second bounded observable. If $b\in M'$, then by the definition of the commutant, \begin{align*} bm=mb \end{align*} for every $m\in M$. Since $E_a(B)\in M$, this gives \begin{align*} bE_a(B)=E_a(B)b \end{align*} for every Borel set $B\subseteq\mathbb R$. Hence every spectral event of $a$ is compatible with $b$. Conversely, suppose $b$ commutes with the whole measurement algebra generated by $a$, meaning \begin{align*} bm=mb \end{align*} for every $m\in M$. This is exactly the condition $b\in M'$ by definition. In particular, because $a\in M$, it also implies \begin{align*} ba=ab. \end{align*} Therefore the bounded observables compatible with all measurement events generated by $a$ are precisely the self-adjoint elements of $M'$. The measurement algebra of one observable is thus controlled by its spectral projections, while compatibility with that algebra is encoded by the commutant $M'$. [/example] The example shows that generated observable algebras are determined by compatibility with the commutant. This motivates the double commutant formulation, which identifies the algebra generated by a set of observables with the observables compatible with every observable compatible with the set. [quotetheorem:1246] [citeproof:1246] This is the measurement version of the bicommutant theorem from earlier chapters. The self-adjoint hypothesis ensures that the generated algebra is closed under adjoints, so it can represent observables rather than just a non-self-adjoint operator algebra; the unit is included so the resulting algebra represents a system with the identity observable. Closure is essential: the polynomial *-algebra generated by one self-adjoint operator contains only continuous functional-calculus approximants, while the von Neumann algebra also contains spectral projections obtained as weak limits. The theorem does not say that the norm-closed C*-algebra generated by $S$ is already $S''$; the point is precisely the extra weak operator, or equivalently ultraweak, closure. The next definition isolates what happens when compatibility is internal to the algebra itself through central projections. [definition: Superselection Projection] Let $M\subseteq\mathcal{L}(H)$ be a von Neumann algebra. A superselection projection for $M$ is a nonzero projection $z\in Z(M)=M\cap M'$. [/definition] Because $z$ commutes with every observable in $M$, no observable in $M$ connects $zH$ with $(1-z)H$. The decomposition by central projections is therefore an algebraic version of a decomposition into sectors, as the following direct-sum model illustrates. [example: Direct Sum Sectors] Let $H=H_1\oplus H_2$, so every vector $\xi\in H$ has a unique form $\xi=\xi_1\oplus \xi_2$ with $\xi_j\in H_j$. Let \begin{align*} M=\mathcal{L}(H_1)\oplus\mathcal{L}(H_2)\subseteq\mathcal{L}(H). \end{align*} Thus an element $a\in M$ has the form $a=a_1\oplus a_2$, where $a_1\in\mathcal{L}(H_1)$ and $a_2\in\mathcal{L}(H_2)$, and it acts by \begin{align*} (a_1\oplus a_2)(\xi_1\oplus \xi_2)=a_1\xi_1\oplus a_2\xi_2. \end{align*} Let $z$ be the projection onto $H_1$, so \begin{align*} z(\xi_1\oplus \xi_2)=\xi_1\oplus 0. \end{align*} Equivalently, $z=1_{H_1}\oplus 0_{H_2}$, hence $z\in M$. Also, \begin{align*} z^2(\xi_1\oplus \xi_2)=z(\xi_1\oplus 0)=\xi_1\oplus 0=z(\xi_1\oplus \xi_2), \end{align*} so $z^2=z$, and $z=z^*$ because $H_1$ is orthogonal to $H_2$ in the direct-sum Hilbert space. Now take $a=a_1\oplus a_2\in M$. On a vector $\xi_1\oplus \xi_2$, \begin{align*} za(\xi_1\oplus \xi_2)=z(a_1\xi_1\oplus a_2\xi_2)=a_1\xi_1\oplus 0. \end{align*} On the other hand, \begin{align*} az(\xi_1\oplus \xi_2)=a(\xi_1\oplus 0)=a_1\xi_1\oplus 0. \end{align*} Thus $za(\xi_1\oplus \xi_2)=az(\xi_1\oplus \xi_2)$ for every $\xi_1\oplus \xi_2\in H$, so $za=az$. Since this holds for every $a\in M$, we have $z\in M'$. Therefore $z\in M\cap M'=Z(M)$, so $z$ is a superselection projection. Finally, each $a=a_1\oplus a_2\in M$ preserves the two summands: \begin{align*} a(\xi_1\oplus 0)=a_1\xi_1\oplus 0\in H_1\oplus 0 \end{align*} and \begin{align*} a(0\oplus \xi_2)=0\oplus a_2\xi_2\in 0\oplus H_2. \end{align*} Thus no observable in $M$ sends a vector from $H_1$ into $H_2$ or from $H_2$ into $H_1$. The central projection $z$ separates the representation into the two superselection sectors $H_1$ and $H_2$. [/example] ## Composite Systems and Independence How should the algebra of a joint quantum system be built from two component systems? There are two related but different notions of independence: tensor product independence, where the two systems sit in separate tensor factors, and commutation independence, where two subalgebras commute inside a larger algebra. The strongest standard construction of a composite quantum system places its two observable algebras on tensor factors. This motivates the spatial tensor product of von Neumann algebras. [definition: Tensor Product Composite System] Let $M\subseteq\mathcal{L}(H)$ and $N\subseteq\mathcal{L}(K)$ be von Neumann algebras. Their spatial tensor product $M\overline{\otimes}N$ is the von Neumann algebra on $H\otimes K$ generated by operators $a\otimes b$ with $a\in M$ and $b\in N$. [/definition] The two subsystems embed as $M\otimes 1$ and $1\otimes N$, and these subalgebras commute. Product states express statistical independence at the level of expectation values, as the next example records. [example: Product State Expectations] Let $\phi$ be a normal state on $M$ and $\psi$ a normal state on $N$. The product state $\phi\otimes\psi$ on $M\overline{\otimes}N$ is characterized on elementary tensors by \begin{align*} (\phi\otimes\psi)(a\otimes b)=\phi(a)\psi(b) \end{align*} for $a\in M$ and $b\in N$. Write the units of $M$ and $N$ as $1_M$ and $1_N$. For elementary vectors $\xi\otimes\eta\in H\otimes K$, \begin{align*} ((a\otimes 1_N)(1_M\otimes b))(\xi\otimes\eta)=(a\otimes 1_N)(\xi\otimes b\eta) \end{align*} and then \begin{align*} (a\otimes 1_N)(\xi\otimes b\eta)=a\xi\otimes b\eta. \end{align*} Also, \begin{align*} (a\otimes b)(\xi\otimes\eta)=a\xi\otimes b\eta. \end{align*} Since elementary tensors span a dense subspace of $H\otimes K$ and both operators are bounded, the two operators are equal: \begin{align*} (a\otimes 1_N)(1_M\otimes b)=a\otimes b. \end{align*} Applying the product state gives \begin{align*} (\phi\otimes\psi)((a\otimes 1_N)(1_M\otimes b))=(\phi\otimes\psi)(a\otimes b). \end{align*} By the defining formula for the product state, \begin{align*} (\phi\otimes\psi)(a\otimes b)=\phi(a)\psi(b). \end{align*} The separate local expectations are \begin{align*} (\phi\otimes\psi)(a\otimes 1_N)=\phi(a)\psi(1_N)=\phi(a) \end{align*} and \begin{align*} (\phi\otimes\psi)(1_M\otimes b)=\phi(1_M)\psi(b)=\psi(b), \end{align*} because states are normalized. Therefore \begin{align*} (\phi\otimes\psi)((a\otimes 1_N)(1_M\otimes b))=(\phi\otimes\psi)(a\otimes 1_N)(\phi\otimes\psi)(1_M\otimes b). \end{align*} Thus tensor product structure, together with a product state, makes expectations of local product observables factor into the expectations of the two subsystem observables. [/example] The product-state example contributes a factorization criterion that is stronger than mere compatibility of measurements. In many von Neumann algebraic settings, one only knows that two subalgebras commute inside a common representation, so this weaker relation needs its own definition. [definition: Commutation Independence] Let $A,B\subseteq\mathcal{L}(H)$ be von Neumann subalgebras. The pair $(A,B)$ is commutation independent in $\mathcal{L}(H)$ if $A\subseteq B'$ and $B\subseteq A'$. [/definition] Tensor product subsystems are commutation independent, but the converse requires extra hypotheses. The failure of the converse is part of what makes von Neumann algebraic quantum theory richer than finite-dimensional matrix mechanics. [example: Bell-Type Separation] Take $H_A=H_B=\mathbb C^2$ with orthonormal basis $e_0,e_1$, and put $H=H_A\otimes H_B$. Let $A=M_2(\mathbb C)\otimes 1$ and $B=1\otimes M_2(\mathbb C)$. For $x,y\in M_2(\mathbb C)$ and an elementary tensor $\xi\otimes\eta$, \begin{align*} ((x\otimes 1)(1\otimes y))(\xi\otimes\eta)=x\xi\otimes y\eta. \end{align*} Similarly, \begin{align*} ((1\otimes y)(x\otimes 1))(\xi\otimes\eta)=x\xi\otimes y\eta. \end{align*} Elementary tensors span $H$, so the two bounded operators are equal: \begin{align*} (x\otimes 1)(1\otimes y)=(1\otimes y)(x\otimes 1). \end{align*} Thus the local algebras commute. Now let $p\in M_2(\mathbb C)$ be the projection onto $\mathbb C e_0$, so $pe_0=e_0$ and $pe_1=0$. For the entangled unit vector \begin{align*} \Omega=\frac{1}{\sqrt 2}(e_0\otimes e_0+e_1\otimes e_1), \end{align*} define the vector state $\omega(T)=(T\Omega,\Omega)$. We compute the two local probabilities. First, \begin{align*} (p\otimes 1)\Omega=\frac{1}{\sqrt 2}e_0\otimes e_0. \end{align*} Since $e_0\otimes e_0$ and $e_1\otimes e_1$ are orthonormal, \begin{align*} \omega(p\otimes 1)=\left(\frac{1}{\sqrt 2}e_0\otimes e_0,\frac{1}{\sqrt 2}(e_0\otimes e_0+e_1\otimes e_1)\right)=\frac12. \end{align*} The same calculation in the second tensor factor gives \begin{align*} \omega(1\otimes p)=\frac12. \end{align*} For the joint local observable, \begin{align*} (p\otimes 1)(1\otimes p)=p\otimes p. \end{align*} Applying it to $\Omega$ gives \begin{align*} (p\otimes p)\Omega=\frac{1}{\sqrt 2}e_0\otimes e_0. \end{align*} Therefore \begin{align*} \omega((p\otimes 1)(1\otimes p))=\omega(p\otimes p)=\frac12. \end{align*} But the product of the separate expectations is \begin{align*} \omega(p\otimes 1)\omega(1\otimes p)=\frac12\cdot\frac12=\frac14. \end{align*} So the commuting local algebras $A$ and $B$ express compatibility of the two measurements, while the entangled state $\omega$ shows that compatibility alone does not force factorization of expectations. [/example] The chapter closes the bridge from abstract von Neumann algebras to quantum mechanics. Observables are represented by self-adjoint affiliated operators, states by normal positive functionals, measurement compatibility by commutants, and generated physical algebras by double commutants. These interpretations will be used later when factors, traces, and type classification are read as structural features of quantum systems rather than as purely algebraic invariants. The quantum-mechanical interpretation now provides the conceptual setting, but inclusions of algebras require a finer tool than states alone. The next chapter studies conditional expectations and subfactors, showing how one averages from a larger algebra onto a smaller one while preserving positivity and normality. # 11. Conditional Expectations and Subfactors In Chapters 3--10, von Neumann algebras were studied through projections, traces, modular structure, and the distinction between types. This chapter studies inclusions $N \subset M$ and asks when the larger algebra can be averaged back onto the smaller one without losing positivity or normality. Conditional expectations are the operator-algebraic analogue of conditional expectation in probability, while subfactors provide a setting where the size of an inclusion becomes a numerical invariant. ## Noncommutative Averaging onto a Subalgebra Suppose $N \subset M$ are von Neumann algebras acting on a Hilbert space $H$. The guiding question is whether every operator in $M$ has a canonical component lying in $N$, in a way compatible with multiplication by elements already in $N$. In a commutative probability algebra this operation is conditional expectation onto a smaller $\sigma$-algebra; in the noncommutative setting, positivity and bimodularity replace pointwise averaging. [definition: Conditional Expectation] Let $N \subset M$ be von Neumann algebras. A conditional expectation from $M$ onto $N$ is a linear map $E:M\to N$ such that: 1. $E$ is normal. 2. $E$ is positive. 3. $E(1_M)=1_N$. 4. $E(x)=x$ for every $x\in N$. 5. $E(a x b)=aE(x)b$ for every $a,b\in N$ and $x\in M$. [/definition] The bimodule identity is the algebraic expression of averaging only over directions transverse to $N$. Once an element is already in $N$, multiplying before or after averaging should not change the averaging operation. The positivity and bimodularity requirements exclude many linear projections that look algebraically harmless. For instance, the linear projection $P:M_2(\mathbb C)\to D_2$ given by \begin{align*} P(ae_{11}+be_{12}+ce_{21}+de_{22})=(a+b)e_{11}+de_{22} \end{align*} fixes the diagonal algebra $D_2$, but it is not positive. The positive rank-one element \begin{align*} \frac{1}{4}e_{11}-\frac{1}{2}e_{12}-\frac{1}{2}e_{21}+e_{22} \end{align*} is sent to \begin{align*} -\frac{1}{4}e_{11}+e_{22}, \end{align*} which is not positive. Positivity is therefore not a cosmetic extra; it is what keeps averaging compatible with the order structure. [example: Diagonal Expectation in Matrices] Let $M=M_n(\mathbb C)$ and let $D_n\subset M_n(\mathbb C)$ be the diagonal subalgebra. For $x=(a_{ij})_{i,j=1}^{n}$, define \begin{align*} E(x)=\operatorname{diag}(a_{11},\dots,a_{nn}). \end{align*} The map is linear because each diagonal entry $a_{ii}$ is a linear coordinate function of $x$, and it is normal because every linear map between finite-dimensional von Neumann algebras is normal. Also, \begin{align*} E(1_n)=\operatorname{diag}(1,\dots,1)=1_n. \end{align*} If $d=\operatorname{diag}(\lambda_1,\dots,\lambda_n)\in D_n$, then \begin{align*} E(d)=\operatorname{diag}(\lambda_1,\dots,\lambda_n)=d, \end{align*} so $E$ fixes $D_n$. To check positivity, let $x\ge 0$ in $M_n(\mathbb C)$. For the standard basis vector $e_i\in\mathbb C^n$, \begin{align*} a_{ii}=\langle xe_i,e_i\rangle\ge 0. \end{align*} Hence $E(x)$ is a diagonal matrix with nonnegative diagonal entries, so $E(x)\ge 0$. Finally, take \begin{align*} d_1=\operatorname{diag}(\lambda_1,\dots,\lambda_n),\qquad d_2=\operatorname{diag}(\mu_1,\dots,\mu_n),\qquad x=(a_{ij}). \end{align*} The $(i,j)$-entry of $d_1xd_2$ is $\lambda_i a_{ij}\mu_j$, so \begin{align*} E(d_1xd_2)=\operatorname{diag}(\lambda_1a_{11}\mu_1,\dots,\lambda_na_{nn}\mu_n). \end{align*} On the other hand, \begin{align*} d_1E(x)d_2=\operatorname{diag}(\lambda_1,\dots,\lambda_n)\operatorname{diag}(a_{11},\dots,a_{nn})\operatorname{diag}(\mu_1,\dots,\mu_n)=\operatorname{diag}(\lambda_1a_{11}\mu_1,\dots,\lambda_na_{nn}\mu_n). \end{align*} Thus $E(d_1xd_2)=d_1E(x)d_2$ for all $d_1,d_2\in D_n$. This conditional expectation keeps exactly the diagonal part of a matrix and removes the matrix coefficients invisible from the diagonal subalgebra. [/example] This example shows the intended geometry: $E$ keeps the part of a matrix visible from the subalgebra and erases the rest. The next structural point is that a positive unital projection onto a C*-subalgebra automatically has strong norm-control properties. [quotetheorem:9310] [citeproof:9310] Tomiyama's theorem explains why conditional expectations are often introduced as norm-one positive projections: the bimodule law is then a consequence rather than an extra axiom. The hypotheses are essential. If positivity is removed, there are unital projections onto C*-subalgebras that are not contractive: for example, \begin{align*} Q(ae_{11}+be_{12}+ce_{21}+de_{22})=(a+b)e_{11}+de_{22} \end{align*} is a unital projection $M_2(\mathbb C)\to D_2$, but it sends the positive rank-one element \begin{align*} \frac{1}{4}e_{11}-\frac{1}{2}e_{12}-\frac{1}{2}e_{21}+e_{22} \end{align*} to \begin{align*} -\frac{1}{4}e_{11}+e_{22}, \end{align*} which is not positive. It also fails bimodularity: with $a=e_{12}$, one has $Q(ae_{11})=0$ while $Q(a)e_{11}=e_{11}$. If the range is not a C*-subalgebra, even positive contractive projections can have no multiplication law to respect, because there is no range multiplication for a bimodule identity to preserve. The theorem also says nothing about whether such a projection exists for a prescribed subalgebra, and in the von Neumann algebra setting it does not by itself give normality. The next result supplies an existence criterion in terms of the modular dynamics attached to a weight. A weight on a von Neumann algebra $M$ is a function $\varphi:M_+\to[0,\infty]$ that is additive and positively homogeneous on the positive cone. It is faithful if $\varphi(x)=0$ for $x\in M_+$ forces $x=0$, normal if it respects suprema of increasing bounded nets of positive operators, and semifinite if every nonzero positive element dominates a nonzero positive element of finite weight. States and semifinite traces are special cases or close relatives of this language; weights are needed here because type III algebras generally do not carry faithful normal semifinite traces. For a faithful normal semifinite weight $\varphi$, Tomita-Takesaki theory associates a modular automorphism group $(\sigma_t^\varphi)_{t\in\mathbb R}$ of $M$. This is a one-parameter group of $*$-automorphisms measuring the nontracial time evolution determined by $\varphi$. With this vocabulary fixed, the condition $\sigma_t^\varphi(N)=N$ below means that the subalgebra $N$ is stable under the modular dynamics of the ambient weight. [quotetheorem:9311] This theorem is stated here as the modular existence criterion for expectations. The proof belongs to the modular theory developed after Tomita-Takesaki theory: invariance of $N$ under $\sigma^\varphi$ allows the standard form and modular objects of $N$ to sit compatibly inside those of $M$, while semifiniteness of $\varphi|_N$ supplies the weight on the target algebra. The conditions are not cosmetic. If a faithful normal state $\varphi$ moves $N$ under its modular automorphism group, then a $\varphi$-preserving normal expectation onto $N$ cannot exist, because such an expectation would force $N$ to be invariant under $\sigma^\varphi$. If $\varphi|_N$ is not semifinite, the target algebra has too few finite-weight positive elements for the $\varphi$-preserving formula to define a semifinite weight on $N$, so the standard Takesaki construction cannot start. The theorem is also weight-dependent: changing the chosen faithful normal semifinite weight changes the modular flow and therefore changes the visible obstruction. [remark: Modular Invariance as the Obstruction] The condition $\sigma_t^\varphi(N)=N$ says that the dynamics determined by the weight $\varphi$ do not move the subalgebra outside itself. In probabilistic language, the state or weight supplies the measure with respect to which averaging is performed, and the restricted weight $\varphi|_N$ must still be semifinite on the target algebra. Takesaki's theorem says that, for faithful normal semifinite weights with semifinite restriction to $N$, modular invariance is the precise compatibility condition needed for a $\varphi$-preserving expectation. [/remark] The finite trace case is the main setting for the rest of this chapter. There the modular group is absent for the trace, and expectations can be built by Hilbert space projection. This contrast is important: outside the finite tracial setting, an orthogonal projection in a Hilbert space representation need not produce a bounded normal map $M\to N$, so the positive construction below is a special feature of finite von Neumann algebras with trace. ## Trace-Preserving Expectations in Finite Algebras The finite case asks for an averaging map that preserves the trace, just as classical conditional expectation preserves integration. The trace also turns $M$ into a Hilbert space after completion, and this allows the expectation to be described as an orthogonal projection. [definition: Finite Von Neumann Algebra with Trace] A finite von Neumann algebra with trace is a pair $(M,\tau)$ where $M$ is a von Neumann algebra and $\tau:M\to\mathbb C$ is a faithful normal tracial state. [/definition] The trace gives the inner product \begin{align*} (x,y)_{L^2(M,\tau)}=\tau(y^*x), \end{align*} and the completion is denoted $L^2(M,\tau)$. This Hilbert space picture raises the key finite question: if $N\subset M$, does orthogonal projection onto $L^2(N,\tau|_N)$ come from an elementwise normal map $M\to N$? [quotetheorem:9312] [citeproof:9312] The finite trace turns conditional expectation into a geometric projection problem, while the algebra structure keeps the projected vector inside a von Neumann subalgebra rather than an arbitrary closed subspace. This viewpoint is most useful when it gives a recognizable criterion for the expected value: elements of $N$ should see $x$ and $E_N(x)$ in the same way through the trace pairing. The next characterization packages that criterion in a form that can be used in examples. It identifies the expectation by its trace pairings against test elements of $N$, so later computations can check the defining relation directly instead of rebuilding the projection each time. [quotetheorem:6904] [citeproof:6904] This criterion is the finite tracial analogue of the classical rule that conditional expectation preserves integrals against all bounded functions measurable with respect to the smaller sigma-algebra. Its scope is deliberately narrow: the trace, common unit, and von Neumann subalgebra hypotheses are part of the setting in which the pairing characterizes the normal trace-preserving expectation. Within that setting it gives a compact computational test. For diagonal matrices it recovers deletion of off-diagonal entries, and for finite group actions it becomes averaging over the group. [example: Fixed-Point Algebra of a Finite Group Action] Let $G$ be a finite group acting on a finite von Neumann algebra $(M,\tau)$ by trace-preserving normal $*$-automorphisms $\alpha_g$, and set \begin{align*} M^G=\{x\in M:\alpha_g(x)=x \text{ for all } g\in G\}. \end{align*} Define \begin{align*} E(x)=\frac{1}{|G|}\sum_{g\in G}\alpha_g(x). \end{align*} For $h\in G$, \begin{align*} \alpha_h(E(x))=\frac{1}{|G|}\sum_{g\in G}\alpha_h(\alpha_g(x))=\frac{1}{|G|}\sum_{g\in G}\alpha_{hg}(x). \end{align*} As $g$ runs through $G$, the element $k=hg$ also runs through $G$, so \begin{align*} \frac{1}{|G|}\sum_{g\in G}\alpha_{hg}(x)=\frac{1}{|G|}\sum_{k\in G}\alpha_k(x)=E(x). \end{align*} Thus $E(x)\in M^G$. The map $E$ is linear and normal because it is a finite average of linear normal maps. It is unital since \begin{align*} E(1)=\frac{1}{|G|}\sum_{g\in G}\alpha_g(1)=\frac{1}{|G|}\sum_{g\in G}1=1. \end{align*} If $x\ge 0$, then each $\alpha_g(x)\ge 0$ because $\alpha_g$ is a $*$-automorphism, and a finite average of positive elements is positive; hence $E(x)\ge 0$. If $x\in M^G$, then $\alpha_g(x)=x$ for every $g\in G$, so \begin{align*} E(x)=\frac{1}{|G|}\sum_{g\in G}x=x. \end{align*} Trace preservation follows from trace preservation of each $\alpha_g$: \begin{align*} \tau(E(x))=\tau\left(\frac{1}{|G|}\sum_{g\in G}\alpha_g(x)\right)=\frac{1}{|G|}\sum_{g\in G}\tau(\alpha_g(x))=\frac{1}{|G|}\sum_{g\in G}\tau(x)=\tau(x). \end{align*} Finally, let $a,b\in M^G$ and $x\in M$. Since $\alpha_g$ is multiplicative and $\alpha_g(a)=a$, $\alpha_g(b)=b$, we have \begin{align*} \alpha_g(axb)=\alpha_g(a)\alpha_g(x)\alpha_g(b)=a\alpha_g(x)b. \end{align*} Therefore \begin{align*} E(axb)=\frac{1}{|G|}\sum_{g\in G}\alpha_g(axb)=\frac{1}{|G|}\sum_{g\in G}a\alpha_g(x)b=a\left(\frac{1}{|G|}\sum_{g\in G}\alpha_g(x)\right)b=aE(x)b. \end{align*} So $E$ is the trace-preserving conditional expectation from $M$ onto $M^G$; it keeps exactly the part of an operator invariant under the [group action](/page/Group%20Action). [/example] Finite group averaging is the prototype for many expectations: a symmetry group determines the subalgebra of invariant observables, and the expectation discards the nonsymmetric part. The next example places this inside factors, where subalgebras have no central decomposition to hide behind. [example: Basic Finite Factor Inclusion] Let $N\subset M$ be an inclusion of finite factors with $1_N=1_M$, and let $\tau$ be the unique normalized trace on $M$. The restriction $\tau|_N$ is normal, faithful, and tracial because these properties are inherited by restriction, and it is normalized since \begin{align*} \tau|_N(1_N)=\tau(1_M)=1. \end{align*} Since $N$ is a finite factor, its normalized trace is unique, so $\tau|_N$ is the normalized trace on $N$. By *Existence of Trace-Preserving Expectation*, there is a unique normal conditional expectation $E_N:M\to N$ satisfying \begin{align*} \tau(E_N(x))=\tau(x) \end{align*} for every $x\in M$. By *Orthogonality Characterization of Trace Expectation*, this map is equivalently characterized by the identities \begin{align*} \tau(y^*E_N(x))=\tau(y^*x)\quad\text{for every }y\in N. \end{align*} These identities determine $E_N(x)$ uniquely: if $z,z'\in N$ both satisfy \begin{align*} \tau(y^*z)=\tau(y^*x)\quad\text{and}\quad \tau(y^*z')=\tau(y^*x)\quad\text{for every }y\in N, \end{align*} then subtracting gives \begin{align*} \tau(y^*(z-z'))=0\quad\text{for every }y\in N. \end{align*} Taking $y=z-z'$ yields \begin{align*} \tau((z-z')^*(z-z'))=0, \end{align*} and faithfulness of $\tau|_N$ gives $z-z'=0$. Thus $E_N(x)$ is the unique element of $N$ with the same $N$-moments as $x$, making it the basic averaging map used to compare $M$ with the subfactor $N$. [/example] The existence of $E_N$ lets us measure the inclusion by asking how small $E_N(x^*x)$ can be compared with $x^*x$. This question leads to the first numerical invariant of a subfactor. ## Jones Index and the Size of a Subfactor For finite-dimensional vector spaces, the dimension ratio $\dim M/\dim N$ measures how large an inclusion is. For factors, ordinary [vector space](/page/Vector%20Space) dimension is not the right invariant, but the trace-preserving conditional expectation still carries a quantitative memory of the inclusion. Jones index packages this size into a number that is stable under isomorphism of inclusions. [definition: Finite Index Subfactor] Let $N\subset M$ be an inclusion of finite factors with common unit and trace-preserving conditional expectation $E_N:M\to N$. The inclusion has finite index if there exists a constant $C>0$ such that \begin{align*} E_N(x^*x)\ge Cx^*x \end{align*} for every $x\in M$. [/definition] The inequality is an order inequality in the positive cone of $M$. It gives a qualitative finite-index condition, but different inclusions can satisfy it with different strengths, and the qualitative statement alone does not measure the size of the inclusion. The next invariant extracts the best possible constant from this order estimate. Because a smaller best lower bound means a larger inclusion, the Jones index is defined as the reciprocal of the optimal Pimsner-Popa constant. [definition: Jones Index via the Pimsner-Popa Constant] Let $N\subset M$ be a finite-index inclusion of finite factors. Its Jones index is \begin{align*} [M:N]=\left(\sup\{C>0:E_N(x^*x)\ge Cx^*x \text{ for all } x\in M\}\right)^{-1}. \end{align*} [/definition] This definition is tailored to conditional expectations and agrees with the usual Jones index for finite-index subfactors. The reciprocal appears because larger inclusions have smaller best lower bound for the expectation. [quotetheorem:9313] In this course the [Pimsner-Popa inequality](/theorems/9313) is used as the first operational form of finite index. The finite-index assumption is essential: for the infinite-index inclusion $\mathbb C1\subset M$ inside a diffuse II$_1$ factor, the trace expectation satisfies $E(p)=\tau(p)1$ on projections, and projections with arbitrarily small trace force every uniform lower bound $C$ to satisfy $C\le \tau(p)$ for arbitrarily small positive numbers. The factor hypothesis is also part of the subfactor statement, not a decoration: for $D_2\subset M_2(\mathbb C)$, the diagonal algebra has central projections, so the same order estimate decomposes over central summands rather than producing a single subfactor index. The trace-preserving expectation matters as well. For $\mathbb C1\subset M_2(\mathbb C)$, the normalized trace expectation gives the usual finite-dimensional index, while a nontracial state projection $x\mapsto \phi(x)1$ can change the best lower bound and, if $\phi$ is a vector state, can make it vanish on a nonzero projection. The estimate is therefore tied to the trace-preserving subfactor expectation, not to an arbitrary positive projection. It does not by itself construct finite-index inclusions or classify which subfactors occur; it assumes the relevant trace-preserving expectation and turns the resulting index into an order estimate. That estimate controls how the inclusion behaves on Hilbert space completions and bimodules. [example: Diagonal Expectation as a Finite-Dimensional Model] Although $D_n\subset M_n(\mathbb C)$ is not a subfactor inclusion because $D_n$ has nontrivial centre, it still illustrates the strength of expectation estimates. In $M_2(\mathbb C)$, take \begin{align*} x_\varepsilon=e_{11}+e_{22}+(1-\varepsilon)(e_{12}+e_{21}),\qquad 0<\varepsilon<1. \end{align*} Its diagonal entries are both $1$, and its off-diagonal entries are both $1-\varepsilon$. For a vector $(s,t)\in\mathbb C^2$, multiplication by $x_\varepsilon$ gives first coordinate $s+(1-\varepsilon)t$ and second coordinate $(1-\varepsilon)s+t$. Hence \begin{align*} \left\langle x_\varepsilon(s,t),(s,t)\right\rangle=|s|^2+|t|^2+(1-\varepsilon)t\overline{s}+(1-\varepsilon)s\overline{t}. \end{align*} Since $t\overline{s}+s\overline{t}=2\operatorname{Re}(s\overline{t})$, this is \begin{align*} |s|^2+|t|^2+2(1-\varepsilon)\operatorname{Re}(s\overline{t}). \end{align*} Using $\operatorname{Re}(s\overline{t})\ge -|s||t|$, we obtain \begin{align*} |s|^2+|t|^2+2(1-\varepsilon)\operatorname{Re}(s\overline{t})\ge |s|^2+|t|^2-2(1-\varepsilon)|s||t|. \end{align*} The right-hand side equals \begin{align*} (|s|-|t|)^2+2\varepsilon |s||t|. \end{align*} This is nonnegative, so $x_\varepsilon\ge 0$. The diagonal expectation deletes the off-diagonal entries, hence \begin{align*} E(x_\varepsilon)=e_{11}+e_{22}=1. \end{align*} To compute the eigenvalues, first apply $x_\varepsilon$ to $(1,1)$: \begin{align*} x_\varepsilon(1,1)=(2-\varepsilon,2-\varepsilon)=(2-\varepsilon)(1,1). \end{align*} Next apply $x_\varepsilon$ to $(1,-1)$: \begin{align*} x_\varepsilon(1,-1)=(\varepsilon,-\varepsilon)=\varepsilon(1,-1). \end{align*} The vectors $(1,1)$ and $(1,-1)$ are linearly independent, so they form a basis of $\mathbb C^2$. Therefore the eigenvalues are $\varepsilon$ and $2-\varepsilon$, and the largest eigenvalue is $2-\varepsilon$. If an order inequality $E(x_\varepsilon)\ge Cx_\varepsilon$ held for this matrix, then testing on the eigenvector $(1,1)$ would give \begin{align*} \left\langle E(x_\varepsilon)(1,1),(1,1)\right\rangle\ge C\left\langle x_\varepsilon(1,1),(1,1)\right\rangle. \end{align*} The left side is \begin{align*} \left\langle (1,1),(1,1)\right\rangle=2. \end{align*} The right side is \begin{align*} C\left\langle (2-\varepsilon)(1,1),(1,1)\right\rangle=2C(2-\varepsilon). \end{align*} Thus $2\ge 2C(2-\varepsilon)$, and dividing by the positive number $2(2-\varepsilon)$ gives \begin{align*} C\le (2-\varepsilon)^{-1}. \end{align*} The example shows why comparing $E(x)$ with $x$ from below is a strong requirement: the diagonal part can be fixed while the off-diagonal correlations push the largest eigenvalue close to $2$. [/example] The factor condition removes central summands and makes the index an invariant of a genuine irreducible size comparison. The simplest finite-dimensional subfactor inclusions come from full matrix algebras embedded with tensor factors. [example: Matrix Tensor Subfactor] Let $k\ge r$, let $N=M_k(\mathbb C)\otimes 1_r$, and let $M=M_k(\mathbb C)\otimes M_r(\mathbb C)$, both equipped with normalized traces. Write $\tau_k$ and $\tau_r$ for the normalized traces on $M_k(\mathbb C)$ and $M_r(\mathbb C)$, so the trace on $M$ is $\tau_k\otimes\tau_r$. Since $M_k(\mathbb C)$ and $M_{kr}(\mathbb C)$ have scalar centres, $N\cong M_k(\mathbb C)$ and $M\cong M_{kr}(\mathbb C)$ are finite factors. Define first on elementary tensors \begin{align*} E_N(a\otimes b)=a\otimes \tau_r(b)1_r. \end{align*} Extending linearly gives $E_N=\operatorname{id}_{M_k}\otimes(\tau_r(\,\cdot\,)1_r)$. If $a\otimes 1_r\in N$, then \begin{align*} E_N(a\otimes 1_r)=a\otimes \tau_r(1_r)1_r=a\otimes 1_r, \end{align*} so $E_N$ fixes $N$. It is trace-preserving because, for elementary tensors, \begin{align*} (\tau_k\otimes\tau_r)(E_N(a\otimes b))=(\tau_k\otimes\tau_r)(a\otimes \tau_r(b)1_r)=\tau_k(a)\tau_r(b)\tau_r(1_r)=\tau_k(a)\tau_r(b), \end{align*} which equals $(\tau_k\otimes\tau_r)(a\otimes b)$, and linearity extends this to all of $M$. To see the index constant, identify $M$ with $r\times r$ block matrices over $M_k(\mathbb C)$. For $y\in M_+$, write \begin{align*} y=\sum_{p,q=1}^r y_{pq}\otimes e_{pq}, \end{align*} with $y_{pq}\in M_k(\mathbb C)$. Then \begin{align*} E_N(y)=\frac{1}{r}\sum_{p=1}^r y_{pp}\otimes 1_r. \end{align*} For a vector $\xi=(\xi_1,\dots,\xi_r)\in(\mathbb C^k)^r$, positivity of $y$ gives the Cauchy-Schwarz inequality for the positive sesquilinear form associated to $y$: \begin{align*} |\langle y_{pq}\xi_q,\xi_p\rangle|\le \langle y_{pp}\xi_p,\xi_p\rangle^{1/2}\langle y_{qq}\xi_q,\xi_q\rangle^{1/2}. \end{align*} Hence \begin{align*} \langle y\xi,\xi\rangle=\sum_{p,q=1}^r \langle y_{pq}\xi_q,\xi_p\rangle\le \left(\sum_{p=1}^r \langle y_{pp}\xi_p,\xi_p\rangle^{1/2}\right)^2. \end{align*} Applying the scalar Cauchy-Schwarz inequality to the $r$ nonnegative numbers $\langle y_{pp}\xi_p,\xi_p\rangle^{1/2}$ gives \begin{align*} \langle y\xi,\xi\rangle\le r\sum_{p=1}^r \langle y_{pp}\xi_p,\xi_p\rangle. \end{align*} Since $\sum_{q=1}^r y_{qq}\ge y_{pp}$ for each $p$, we also have \begin{align*} r\sum_{p=1}^r \langle y_{pp}\xi_p,\xi_p\rangle\le r\sum_{p=1}^r \left\langle \left(\sum_{q=1}^r y_{qq}\right)\xi_p,\xi_p\right\rangle. \end{align*} The right side is $\langle r^2E_N(y)\xi,\xi\rangle$, so $y\le r^2E_N(y)$. Taking $y=x^*x$ gives \begin{align*} E_N(x^*x)\ge r^{-2}x^*x. \end{align*} The constant cannot be improved. Choose orthonormal vectors $u_1,\dots,u_r$ in $\mathbb C^k$ and the standard basis $f_1,\dots,f_r$ of $\mathbb C^r$. Put \begin{align*} \eta=\frac{1}{\sqrt r}\sum_{j=1}^r u_j\otimes f_j \end{align*} and let $y=p_\eta$ be the rank-one projection onto $\mathbb C\eta$. Since \begin{align*} p_\eta=\frac{1}{r}\sum_{i,j=1}^r u_i u_j^*\otimes f_i f_j^* \end{align*} and $\tau_r(f_i f_j^*)=\delta_{ij}/r$, we get \begin{align*} E_N(y)=\frac{1}{r^2}\left(\sum_{i=1}^r u_i u_i^*\right)\otimes 1_r. \end{align*} Evaluating on $\eta$ gives $\langle E_N(y)\eta,\eta\rangle=r^{-2}$, while $y\eta=\eta$. Therefore any inequality $E_N(y)\ge Cy$ forces $C\le r^{-2}$. Together with the lower bound above, the best reciprocal index constant is $r^{-2}$, hence \begin{align*} [M:N]=r^2. \end{align*} The inclusion adds the full $M_r(\mathbb C)$ tensor direction, and the index records the square of the Hilbert-space dimension of that added matrix factor. [/example] This example is the finite-dimensional shadow of the general theory: the inclusion adds an $M_r(\mathbb C)$ tensor direction, and the index records the square of its Hilbert space dimension. In infinite-dimensional II$_1$ factors the same formalism survives even when no literal tensor decomposition is present. [remark: Why Index Belongs with Conditional Expectation] The index is not first defined by counting basis vectors inside $M$. It is detected by the best lower bound for the conditional expectation $E_N$, so it is an analytic invariant of the averaging operation. This viewpoint is essential in the general theory, where finite index inclusions are studied through $N$-$N$ bimodules, Jones projections, and the basic construction. [/remark] The chapter therefore moves from existence of averaging maps to quantitative control of inclusions. Conditional expectations provide the bridge between state-preserving structure and subfactor index: they are simultaneously projections, averaging operators, and the analytic device by which one compares $N$ with $M$. Conditional expectations connect inclusion theory to averaging, index, and subfactor structure, but they also complete the bridge between analytic and categorical viewpoints. The final chapter gathers the classification results, representation theory, and modular phenomena into a single picture of how von Neumann algebras are organized and how far the preceding invariants can take the classification problem. # 12. Synthesis: Classification Landscape and Further Directions This final chapter synthesizes the classification theory developed in Chapters 5--9 and connects it with the representation, state, and subfactor viewpoints from Chapters 10--11. The guiding question is how far projections, traces, centres, and modular phenomena take us toward an [isomorphism classification](/theorems/3309) of von Neumann algebras. The answer is deliberately layered: first classify factors, then understand how a nontrivial centre assembles factor-like pieces, and finally recognize where additional invariants beyond type are needed. ## The Type Table as a Classification Map The classification problem begins with a practical question: given a von Neumann algebra $M \subseteq \mathcal{L}(H)$, what features survive isomorphism and allow us to place $M$ into one of the standard types? The answer is not a single invariant, but a hierarchy: first the centre, then the factor summands, then the projection comparison structure inside each factor. [definition: Factor] A von Neumann algebra $M$ is a factor if its centre satisfies \begin{align*} Z(M) := M \cap M' = \mathbb C I. \end{align*} [/definition] A factor is the indecomposable case for central decomposition. To classify a factor we need an internal notion of when two subspaces have the same size, because ordinary dimension is not available in arbitrary von Neumann algebras. [definition: Murray Von Neumann Equivalence] Let $M$ be a von Neumann algebra and let $p,q \in M$ be projections. The projections $p$ and $q$ are Murray-von Neumann equivalent, written $p \sim q$, if there exists a partial isometry $v \in M$ such that \begin{align*} v^*v = p, \qquad vv^* = q. \end{align*} [/definition] This relation replaces rank in settings where there may be no basis and no finite-dimensional dimension function. The next distinction asks whether a projection is the same size as a proper part of itself, since this separates finite matrix-like behavior from genuinely infinite behavior. [definition: Finite And Infinite Projection] Let $M$ be a von Neumann algebra and let $p \in M$ be a projection. The projection $p$ is finite if there is no proper subprojection $q < p$ with $q \sim p$. The projection $p$ is infinite if it is not finite. [/definition] Finiteness is a property of projections relative to the ambient algebra, not just of their Hilbert-space ranges. To sort factors, one must also ask whether minimal projections exist, whether the identity is finite or infinite, and whether trace-theoretic semifiniteness is available. These tests separate matrix algebras, full operator algebras on infinite-dimensional Hilbert spaces, diffuse finite factors, diffuse semifinite infinite factors, and the trace-free purely infinite regime. [definition: Type Of A Factor] Let $M$ be a factor. It is of type $I_n$ if it is isomorphic to $M_n(\mathbb C)$ for $n \in \mathbb N$; of type $I_\infty$ if it is isomorphic to $\mathcal{L}(K)$ for an infinite-dimensional Hilbert space $K$; of type $II_1$ if the identity projection $I$ is finite, $M$ is infinite-dimensional, and $M$ has no nonzero minimal projections; of type $II_\infty$ if $M$ is semifinite, the identity projection $I$ is infinite, and $M$ has no nonzero minimal projections; and of type $III$ if every nonzero projection in $M$ is infinite. [/definition] The table should be read as a decision tree, but the definition itself does not prove that every factor lands in exactly one branch. A factor might a priori fail all listed alternatives, or two of the projection-theoretic tests might overlap in an unexpected way. The classification theorem is needed to turn the table into a theorem rather than terminology. It rules out missing and overlapping cases: the alternatives are exhaustive and mutually exclusive, so the type table is a genuine first invariant of factors rather than a list of examples. [quotetheorem:9314] The hypotheses here are all diagnostic rather than decorative. The factor assumption is essential: $M_2(\mathbb C)\oplus R$ has both type $I_2$ and type $II_1$ central summands, so it cannot belong to a single factor type. Minimal projections separate type I behavior from diffuse type II behavior; for example $\mathcal{L}(H)$ and the hyperfinite $II_1$ factor both have many nonzero projections, but only $\mathcal{L}(H)$ has rank-one minimal projections. Finiteness of the identity separates $II_1$ from $II_\infty$, while the existence of a faithful normal semifinite trace separates semifinite type II behavior from type III behavior. The theorem does not classify factors up to isomorphism inside a fixed type: there are many non-isomorphic type $II_1$ factors and many non-isomorphic type III factors. What it provides is the first sorting map. The rest of the chapter uses this map to explain how central decompositions assemble nonfactor algebras and why finer invariants are needed after the type has been identified. [example: Sorting The Standard Factors] Let $H$ be a separable infinite-dimensional Hilbert space with orthonormal basis $(e_k)_{k\ge 0}$. In $\mathcal{L}(H)$, the centre is $\mathbb C I$, so $\mathcal{L}(H)$ is a factor. The projection $p_0$ onto $\mathbb C e_0$ is minimal: if $0\le q\le p_0$ is a projection, then $qH\subseteq p_0H=\mathbb C e_0$, so $qH$ is either $0$ or $\mathbb C e_0$, hence $q=0$ or $q=p_0$. The identity is infinite because the unilateral shift $S e_k=e_{k+1}$ satisfies \begin{align*} S^*S=I,\qquad SS^*=I-p_0<I. \end{align*} Thus $I\sim I-p_0$ with $I-p_0$ a proper subprojection of $I$, and the type table places $\mathcal{L}(H)$ in type $I_\infty$. For $M_n(\mathbb C)$, the centre is again $\mathbb C I$, and rank-one projections are minimal for the same reason: every subprojection of a one-dimensional range projection has range either $0$ or that same one-dimensional subspace. The identity is finite. Indeed, if $q\le I$ and $q\sim I$, choose $v\in M_n(\mathbb C)$ with \begin{align*} v^*v=I,\qquad vv^*=q. \end{align*} The equation $v^*v=I$ says that $v:\mathbb C^n\to\mathbb C^n$ is an isometry, so it is injective; in finite dimension injective linear maps are surjective, hence $vv^*=I$. Therefore $q=I$, so no proper subprojection is equivalent to $I$. This is the diagnostic separating type $I_n$ from type $I_\infty$. The hyperfinite $II_1$ factor $R$ is finite and diffuse. It has a faithful normal tracial state $\tau$ with $\tau(I)=1$, and it has no nonzero minimal projections. Finiteness follows visibly from the trace: if $q<I$ and $q\sim I$, then for a partial isometry $v$ with $v^*v=I$ and $vv^*=q$, \begin{align*} \tau(q)=\tau(vv^*)=\tau(v^*v)=\tau(I)=1. \end{align*} Hence $\tau(I-q)=\tau(I)-\tau(q)=0$, and faithfulness gives $I-q=0$, contradicting $q<I$. Thus $R$ is type $II_1$, not merely another infinite-dimensional algebra. The amplification $R\overline{\otimes}\mathcal{L}(\ell^2(\mathbb N))$ remains diffuse because the $R$ tensor factor has no nonzero minimal projections, and it is semifinite using the tensor product of the tracial state on $R$ with the usual semifinite trace on $\mathcal{L}(\ell^2(\mathbb N))$. Its identity is infinite: with the unilateral shift $S$ on $\ell^2(\mathbb N)$, \begin{align*} (1_R\otimes S)^*(1_R\otimes S)=1_R\otimes I,\qquad (1_R\otimes S)(1_R\otimes S)^*=1_R\otimes (I-p_0)<1_R\otimes I. \end{align*} So the amplification is type $II_\infty$. Finally, a local algebra in algebraic quantum field theory that is known to be a type III factor has no nonzero finite projections by definition of type III. Therefore no nonzero projection can carry a finite trace-type size compatible with Murray-von Neumann equivalence, and the faithful normal semifinite trace test fails on the whole factor. [/example] The examples show that the type is not just a dimension count. Still, in the model type I algebra $\mathcal{L}(H)$, Murray-von Neumann equivalence ought to recover the familiar comparison of subspaces by Hilbert-space dimension. The point that must be checked is that every Hilbert-space isomorphism between the ranges of two projections is implemented by a partial isometry lying in the full operator algebra, so the abstract equivalence relation agrees with ordinary geometry in this special case. This comparison is the calibration theorem for the whole equivalence relation. Before using Murray-von Neumann equivalence as a replacement for dimension in more general algebras, we should verify that in the full operator algebra it gives exactly the dimension test already familiar from Hilbert space geometry. The precise question is therefore not whether equal-dimensional ranges can be matched abstractly, but whether that matching can be realized by an operator inside $\mathcal{L}(H)$ with the prescribed initial and final projections. The following result records exactly this realization criterion, making the familiar Hilbert-space dimension comparison the first concrete model for Murray-von Neumann equivalence. [quotetheorem:9315] [citeproof:9315] The algebra $\mathcal{L}(H)$ hypothesis is essential. In a proper subalgebra, two projections can have ranges of the same Hilbert-space dimension but fail to be equivalent because the required partial isometry need not lie in the algebra; diagonal projections in an abelian von Neumann algebra give the simplest warning. The theorem also does not say that Hilbert-space dimension classifies projections in arbitrary factors, or that equal trace always gives equivalence without the appropriate comparison theory. Its role is to identify the model case: type I classification feels familiar precisely because Murray-von Neumann equivalence reduces to ordinary Hilbert-space dimension there, while later types require trace-valued dimension functions or modular invariants. ## Central Decomposition and Nonfactor Algebras Most von Neumann algebras encountered in analysis are not factors. A direct sum such as $M_2(\mathbb C)\oplus L^\infty([0,1])$ already shows the problem: one summand is a matrix factor, while the other is abelian and decomposes over a continuum of central projections, so assigning a single factor type to the whole algebra loses information. The next question is therefore how the factor classification extends when the centre is nontrivial. Central projections split the algebra into pieces, and the type classification is applied after this splitting. [definition: Central Projection] Let $M$ be a von Neumann algebra. A projection $z \in M$ is central if $z \in Z(M)$. [/definition] A central projection behaves like a measurable characteristic function for a direct summand. Since factor theory is only the fibrewise picture, the classification of general von Neumann algebras requires a theorem saying that the centre separates the different type regions. [quotetheorem:9316] The centrality hypothesis is what makes the statement possible. Noncentral projections cut out corners, but corners need not describe direct summands of the algebra; central projections are the pieces visible to the centre and therefore behave like measurable characteristic functions of type regions. The theorem does not assert that $M$ is a direct sum of individual factors, nor that the central projections are atoms. For example, an abelian algebra $L^\infty(X,\mu)$ may have a diffuse centre, so the type $I_1$ part is spread over a measure space rather than over a discrete list. The mechanism is to form the largest central projection supporting each projection-theoretic property. One first extracts the central support of all type I behavior, then separates the remaining semifinite continuous part from the residual part with no finite central summand. This is why the theorem is the bridge from factor theory to general algebras: after it, every classification question can be asked fibrewise over the centre, with the warning that the centre itself may be continuous. [example: Abelian Von Neumann Algebras] Let $(X,\mathcal E,\mu)$ be a localizable measure space, and let $M=L^\infty(X,\mu)$ act on $L^2(X,\mu)$ by multiplication. Because multiplication of functions is commutative, for $f,g\in L^\infty(X,\mu)$ we have $M_fM_g=M_{fg}=M_{gf}=M_gM_f$, so every element of $M$ commutes with every element of $M$. Hence $Z(M)=M$. Therefore $M$ is a factor exactly when $M=\mathbb C I$, which means every essentially bounded measurable function is almost everywhere constant; equivalently, the measure algebra has only the zero class and the unit class. The projections in $M$ are exactly multiplication operators by indicator functions. Indeed, if $M_f$ is a projection, then $M_f^*=M_f$ and $M_f^2=M_f$, so $f=\overline f$ almost everywhere and \begin{align*} f(x)^2=f(x)\text{ for almost every }x. \end{align*} Thus $f(x)\in\{0,1\}$ almost everywhere, so $f=\mathbf 1_A$ almost everywhere for a measurable set $A$. Conversely, $M_{\mathbf 1_A}^*=M_{\mathbf 1_A}$ and \begin{align*} M_{\mathbf 1_A}^2=M_{\mathbf 1_A^2}=M_{\mathbf 1_A}, \end{align*} so $M_{\mathbf 1_A}$ is a projection. A central summand is therefore obtained by choosing a measurable set $A$ and cutting by $z=M_{\mathbf 1_A}$; the corner $zM$ is $L^\infty(A,\mu|_A)$ acting on $L^2(A,\mu|_A)$. If this summand is a factor, then $zM=\mathbb C z$. Its only projections are $0$ and $z$, so $z$ is minimal. Also $z$ is finite: if $v\in \mathbb C z$ satisfies $v^*v=z$, then $v=\lambda z$ and \begin{align*} v^*v=|\lambda|^2z=z, \end{align*} so $|\lambda|^2=1$ and hence \begin{align*} vv^*=|\lambda|^2z=z. \end{align*} Thus no proper subprojection of $z$ is Murray-von Neumann equivalent to $z$. Every nonzero abelian factor summand is therefore $\mathbb C$, hence type $I_1$; abelian von Neumann algebras decompose measure-theoretically into type $I_1$ pieces, not into type II or type III factor pieces. [/example] This example is a useful warning: type theory does not say that every von Neumann algebra is close to a single factor. It says that the centre is the parameter space over which factor-like behavior varies. [example: Group Von Neumann Algebras In The Classification Picture] For a discrete group $G$, define the left and right regular representations on the basis $(\delta_h)_{h\in G}$ of $\ell^2(G)$ by \begin{align*} \lambda_g\delta_h=\delta_{gh} \end{align*} and \begin{align*} \rho_k\delta_h=\delta_{hk^{-1}}. \end{align*} Then \begin{align*} \lambda_g\rho_k\delta_h=\lambda_g\delta_{hk^{-1}}=\delta_{ghk^{-1}} \end{align*} and \begin{align*} \rho_k\lambda_g\delta_h=\rho_k\delta_{gh}=\delta_{ghk^{-1}}, \end{align*} so $\lambda_g\rho_k=\rho_k\lambda_g$ for all $g,k\in G$. The group von Neumann algebra $L(G)$ is the weak operator closure of the linear span of the operators $\lambda_g$. To see why ICC groups give factors, first look at a finite linear combination \begin{align*} x=\sum_{g\in F} a_g\lambda_g. \end{align*} For $s\in G$, using $\lambda_s^*=\lambda_{s^{-1}}$ and $\lambda_a\lambda_b=\lambda_{ab}$ gives \begin{align*} \lambda_s x\lambda_s^*=\sum_{g\in F}a_g\lambda_s\lambda_g\lambda_{s^{-1}}=\sum_{g\in F}a_g\lambda_{sgs^{-1}}. \end{align*} Thus centrality forces the coefficient of $\lambda_h$ to equal the coefficient of $\lambda_{s^{-1}hs}$ for every $s,h\in G$; the Fourier coefficients are constant on conjugacy classes. The same coefficient relation holds for general $x\in L(G)$ by applying both sides to $\delta_e$ and comparing the $\ell^2(G)$ coefficients of $x\delta_e$. If $G$ is ICC, every nonidentity conjugacy class is infinite, and an $\ell^2(G)$ function that is constant with value $c$ on an infinite set must have $c=0$, since otherwise the sum of squared absolute values over that class would be infinite. Hence only the coefficient at $e$ can remain, so every central element is a scalar multiple of $I$. Therefore \begin{align*} Z(L(G))=\mathbb C I, \end{align*} and $L(G)$ is a factor. The vector state \begin{align*} \tau(x)=\langle x\delta_e,\delta_e\rangle \end{align*} is the canonical faithful normal trace on $L(G)$. On the group generators it has value $\tau(\lambda_e)=1$, because $\lambda_e=I$, and for $g\ne e$ it has value \begin{align*} \tau(\lambda_g)=\langle \delta_g,\delta_e\rangle=0. \end{align*} So an infinite ICC group gives a finite factor with canonical trace; when the group is also amenable, the Connes hyperfiniteness theorem identifies this factor with the hyperfinite type $II_1$ factor $R$. Thus group von Neumann algebras turn conjugacy and amenability of $G$ into the factor and type labels in the classification table. [/example] Group von Neumann algebras show how algebra and dynamics enter the classification landscape. A group with rich conjugacy behavior can produce a factor, and amenability pushes many such examples toward hyperfinite structure. ## Semifinite Traces and Purely Infinite Behavior Once the algebra has no minimal projections, the next problem is whether projections can still be measured. Semifinite algebras retain enough finite projections to build trace theory, while type III algebras do not. [definition: Faithful Normal Semifinite Trace] Let $M$ be a von Neumann algebra. A map $\tau:M_+\to [0,\infty]$ is a faithful normal semifinite trace if it is additive, positively homogeneous, satisfies $\tau(x^*x)=\tau(xx^*)$ for all $x\in M$, is normal on increasing bounded nets in $M_+$, is faithful in the sense that $\tau(x)=0$ implies $x=0$, and for every nonzero $x\in M_+$ there exists $0\le y\le x$ with $0<\tau(y)<\infty$. [/definition] Semifiniteness says that the algebra has enough finite mass pieces to approximate positive operators from below. This leads to a class of von Neumann algebras where noncommutative integration can be built from such traces. [definition: Semifinite Von Neumann Algebra] A von Neumann algebra $M$ is semifinite if it admits a faithful normal semifinite trace. [/definition] The finite and semifinite types are those where trace methods remain central, but a trace can serve as a dimension function only if it is insensitive to replacing a projection by an equivalent one. Murray-von Neumann equivalence identifies projections through a partial isometry, so the relevant compatibility check is whether the trace gives the same value to the initial and final projections of that partial isometry. Without this invariance, trace values would not measure projection size in the sense used by the type table. The next ingredient is therefore not another definition of trace, but a consistency test between the trace and the projection comparison relation. To justify trace-based dimension theory, one must know that equivalent projections cannot receive different trace sizes. [quotetheorem:9317] [citeproof:9317] The trace hypothesis is essential in two ways. Without the trace property, a positive weight need not take equal values on $v^*v$ and $vv^*$; without semifiniteness, the weight may assign only infinite values on the part one wants to measure. The theorem does not say that equal trace values force Murray-von Neumann equivalence in every von Neumann algebra; that converse requires additional comparison hypotheses and can fail when the centre or projection lattice is too large. What the theorem gives is the compatibility direction needed for classification: any trace-based dimension function must be constant on Murray-von Neumann size classes. The preceding compatibility result explains why semifinite algebras admit a trace-based dimension theory: finite trace pieces survive projection equivalence and can be assembled into noncommutative integration. The opposite regime is not detected by looking only at the identity, because properly infinite semifinite algebras still contain many finite projections. What must be ruled out is a finite direct summand visible to the centre. This leads to the following definition, which describes algebras whose central decomposition has no finite part. [definition: Purely Infinite Von Neumann Algebra] A von Neumann algebra $M$ is purely infinite if there is no nonzero central projection $z\in Z(M)$ such that the von Neumann algebra $zM$ is finite. [/definition] Thus every nonzero central summand of a purely infinite von Neumann algebra lies outside finite trace-type dimension theory. This is a central decomposition condition, not merely a statement that the identity projection is infinite. The obstruction is that trace theory requires enough finite positive pieces to approximate nonzero positive operators, while type III factors have no finite projection structure from which such a dimension theory can start. If a faithful normal semifinite trace existed on a type III factor, its semifinite finite-mass pieces would contradict the projection-theoretic meaning of type III. [quotetheorem:9318] [citeproof:9318] The factor and type III hypotheses are both doing work. A properly infinite semifinite factor such as $\mathcal{L}(H)$ has many infinite projections but still carries the usual faithful normal semifinite trace, so infinite identity alone is not enough to force the conclusion. The theorem also does not say that every algebra without a faithful normal semifinite trace is a type III factor; nonfactor algebras may fail semifiniteness for central-decomposition reasons, and type III is a factor-level label. The forward message is the main conceptual boundary in the table: semifinite algebras are studied with traces, noncommutative $L^p$-spaces, and dimension functions, while type III algebras require modular theory even to state their finer invariants. ## Landmark Classification Results and Examples The final classification question asks how far the type table goes. The table is necessary for orientation, but it is not a complete isomorphism classification inside each type. The deepest results identify special subclasses where the table can be sharpened by additional invariants. The classification theorem below uses three pieces of standard terminology. A factor is injective when it satisfies the operator-algebraic extension property for completely positive maps; for this course, injectivity should be read as the strong regularity hypothesis that makes the factor accessible to hyperfinite and approximation methods. The Connes-Haagerup invariants are the modular and flow-theoretic data used to distinguish injective factors, especially in type III. The flow of weights is a measure-theoretic flow attached to a type III von Neumann algebra that records the residual modular behavior left after ordinary trace-based dimension has disappeared. Thus the theorem is not saying that the coarse type label alone classifies all factors. It says that, within the injective separable-predual setting, the type label together with the relevant Connes-Haagerup modular invariants gives the complete classification. [quotetheorem:9319] The separable predual and injectivity hypotheses are essential. Without injectivity there are many non-isomorphic factors of the same type, including many type $II_1$ factors not isomorphic to $R$; without the standard separability hypothesis, the usual uniqueness statements are not the form used in this classification theorem. The theorem also does not reduce type III classification to a single numerical invariant in all cases: type $III_0$ requires the flow of weights, while the parameters $0<\lambda\le 1$ describe the better-behaved type $III_\lambda$ cases. The mechanism lies far beyond the type table. Injectivity is converted into hyperfinite or approximately finite-dimensional structure, modular theory supplies the flow and Connes spectrum, and crossed-product analysis reconstructs the factor from these data. Thus the result is a guide to the research landscape: the type table is the entrance to classification, while injectivity and modular invariants are what make a complete theorem possible in this special class. [example: Hyperfinite Approximation] Let $A_k=M_{2^k}(\mathbb C)$, and embed $A_k$ into $A_{k+1}$ by \begin{align*} a\longmapsto a\otimes I_2. \end{align*} With normalized traces $\tau_k(a)=2^{-k}\operatorname{Tr}(a)$, this embedding preserves trace because \begin{align*} \tau_{k+1}(a\otimes I_2)=2^{-(k+1)}\operatorname{Tr}(a\otimes I_2)=2^{-(k+1)}\operatorname{Tr}(a)\operatorname{Tr}(I_2)=2^{-k}\operatorname{Tr}(a)=\tau_k(a). \end{align*} Thus the union $A_\infty=\bigcup_{k\ge 1}A_k$ carries a well-defined tracial state $\tau$ by setting $\tau(a)=\tau_k(a)$ whenever $a\in A_k$. Represent $A_\infty$ on the GNS Hilbert space $L^2(A_\infty,\tau)$ by left multiplication, and define \begin{align*} R=\overline{A_\infty}^{\,\mathrm{WOT}}\subseteq \mathcal L(L^2(A_\infty,\tau)). \end{align*} The vector state induced by the GNS cyclic vector extends $\tau$ to a faithful normal tracial state on $R$, because each finite-stage trace is faithful and tracial and the weak closure preserves normality of the induced vector state. The algebra $R$ has no nonzero minimal projections. Indeed, if $p\in A_k$ is a nonzero projection, then inside $A_{k+1}=A_k\otimes M_2(\mathbb C)$ we have \begin{align*} p=(p\otimes e_{11})+(p\otimes e_{22}), \end{align*} where $e_{11},e_{22}$ are the two diagonal rank-one projections in $M_2(\mathbb C)$. Both summands are projections, they are orthogonal, and \begin{align*} p\otimes e_{11}\ne 0,\qquad p\otimes e_{22}\ne 0. \end{align*} So every nonzero finite-stage projection splits into two nonzero smaller projections at the next stage, and the weak closure retains diffuse projection structure rather than producing atoms. The trace is finite on the identity: \begin{align*} \tau(I)=1. \end{align*} Therefore this construction gives a finite diffuse factor with faithful normal tracial state: the hyperfinite $II_1$ factor $R$. Its hyperfiniteness is visible in the construction itself, since $R$ is generated as a weak operator closure by the increasing finite-dimensional matrix algebras $M_{2^k}(\mathbb C)$. [/example] Hyperfiniteness is the operator-algebraic analogue of approximation by finite systems. It is why $R$ appears both in abstract classification and in models arising from statistical mechanics and quantum information. [example: Type III Local Algebras] Let $\mathcal O$ be a bounded spacetime region, and suppose its local observable algebra $A(\mathcal O)$ is known to be a type III factor. Then \begin{align*} Z(A(\mathcal O))=\mathbb C I. \end{align*} Moreover, every nonzero projection $p\in A(\mathcal O)$ is infinite: there is a proper subprojection $q<p$ and a partial isometry $v\in A(\mathcal O)$ such that \begin{align*} v^*v=p,\qquad vv^*=q. \end{align*} Thus no nonzero projection in $A(\mathcal O)$ can behave like a finite-dimensional rank projection. In particular, $A(\mathcal O)$ cannot be type I. If $e$ were a nonzero minimal projection, then the only subprojections of $e$ would be $0$ and $e$. Therefore there could be no proper subprojection $q<e$ with $q\sim e$, so $e$ would be finite. This contradicts the type III condition that every nonzero projection is infinite. Hence $A(\mathcal O)$ has no nonzero minimal projections, unlike $M_n(\mathbb C)$ or $\mathcal L(H)$. This is the operator-algebraic form of the physical distinction: a local subsystem in relativistic quantum field theory is not modeled by a tensor factor with ordinary density-matrix trace theory, because its local algebra has type III projection structure rather than type I finite-rank structure. [/example] The contrast with matrix algebras is sharp. Quantum information often begins with $M_n(\mathbb C)$ and normal states $a\mapsto \operatorname{Tr}(\rho a)$, while quantum field theory naturally leads to local algebras where modular theory is part of the basic structure. [explanation: Diagnostic Checklist For A Concrete Algebra] To classify a concrete von Neumann algebra $M$, first compute or identify the centre $Z(M)$. If $Z(M)=\mathbb C I$, proceed as a factor; if not, look for central projections that separate type regions before assigning any single label. A direct sum with two different summand types is the standard test case showing why this first step cannot be skipped. For each factor summand, ask whether nonzero minimal projections exist. Their presence points to type I, and the remaining distinction is whether the identity has finite rank-like size or infinite rank-like size. If there are no minimal projections, test whether the identity is finite and whether a faithful normal semifinite trace exists. Finite diffuse trace behavior gives type $II_1$, infinite semifinite diffuse behavior gives type $II_\infty$, and the absence of any nonzero finite projection gives type III. Finally, remember what this checklist does not decide. It assigns type, but it does not classify factors up to isomorphism inside type $II_1$ or type III. For that, one needs the finer tools indicated above: hyperfiniteness, injectivity, crossed products, modular invariants, and rigidity phenomena. [/explanation] The checklist separates the infinite-dimensional obstructions from the finite-dimensional model that first motivated traces and states. To close the circle, it is useful to return to matrix algebras and record exactly what the trace picture says there: normal states are concrete density matrices, and this is the model that type II and type III theory generalize or replace. [example: Normal States On Matrix Algebras] Let $M=M_n(\mathbb C)$, and let $(e_{ij})_{1\le i,j\le n}$ be the standard matrix units. We compute the density matrix of a state $\omega:M\to\mathbb C$ explicitly. Define $\rho=(\rho_{ij})$ by \begin{align*} \rho_{ij}=\omega(e_{ji}). \end{align*} If $a=(a_{ij})\in M_n(\mathbb C)$, then \begin{align*} a=\sum_{i=1}^n\sum_{j=1}^n a_{ij}e_{ij}. \end{align*} By linearity of $\omega$, \begin{align*} \omega(a)=\sum_{i=1}^n\sum_{j=1}^n a_{ij}\omega(e_{ij}). \end{align*} On the other hand, the $i$th diagonal entry of $\rho a$ is \begin{align*} (\rho a)_{ii}=\sum_{j=1}^n \rho_{ij}a_{ji}. \end{align*} Therefore \begin{align*} \operatorname{Tr}(\rho a)=\sum_{i=1}^n(\rho a)_{ii}=\sum_{i=1}^n\sum_{j=1}^n\rho_{ij}a_{ji}=\sum_{i=1}^n\sum_{j=1}^n\omega(e_{ji})a_{ji}=\omega(a). \end{align*} The matrix $\rho$ is positive. For $\xi=(\xi_1,\dots,\xi_n)\in\mathbb C^n$, test positivity of $\omega$ on the rank-one matrix \begin{align*} b=\sum_{i=1}^n \overline{\xi_i}e_{1i}. \end{align*} Then \begin{align*} b^*b=\sum_{i=1}^n\sum_{j=1}^n \xi_i\overline{\xi_j}e_{i1}e_{1j}=\sum_{i=1}^n\sum_{j=1}^n \xi_i\overline{\xi_j}e_{ij}. \end{align*} Since $\omega$ is positive, \begin{align*} 0\le \omega(b^*b)=\sum_{i=1}^n\sum_{j=1}^n \xi_i\overline{\xi_j}\omega(e_{ij})=\sum_{i=1}^n\sum_{j=1}^n \overline{\xi_j}\rho_{ji}\xi_i=\langle \rho\xi,\xi\rangle. \end{align*} Thus $\rho\ge 0$. Also \begin{align*} \operatorname{Tr}(\rho)=\operatorname{Tr}(\rho I)=\omega(I)=1. \end{align*} Conversely, if $\rho\ge 0$ and $\operatorname{Tr}(\rho)=1$, then the formula $\omega_\rho(a)=\operatorname{Tr}(\rho a)$ defines a linear functional with $\omega_\rho(I)=1$. If $a\ge 0$, write $a=c^*c$ for some matrix $c$. Then \begin{align*} \omega_\rho(a)=\operatorname{Tr}(\rho c^*c)=\operatorname{Tr}(c\rho c^*). \end{align*} Because $\rho\ge 0$, the matrix $c\rho c^*$ is positive, so its trace is nonnegative. Hence $\omega_\rho$ is a state. Uniqueness follows from the matrix units: if $\operatorname{Tr}(\rho a)=\operatorname{Tr}(\sigma a)$ for all $a$, then taking $a=e_{ij}$ gives \begin{align*} \rho_{ji}=\operatorname{Tr}(\rho e_{ij})=\operatorname{Tr}(\sigma e_{ij})=\sigma_{ji}. \end{align*} So $\rho=\sigma$. Thus normal states on $M_n(\mathbb C)$ are exactly density matrices, and the trace formula is the finite-dimensional model for representing states by positive unit-mass operators. [/example] This finite-dimensional picture is the simplest point in the classification landscape. The rest of the course can be viewed as a systematic account of which parts of this picture survive for general von Neumann algebras and which parts must be replaced. ## Connections Beyond The Course The course ends by asking why von Neumann algebras sit at a crossroads of several fields. The same formalism describes spectral decompositions in operator theory, invariant measurable structures in ergodic theory, states and channels in quantum information, and local observables in quantum field theory. [explanation: Operator Theory And Spectral Analysis] Von Neumann algebras refine the spectral theorem by studying all operators compatible with a family of observables. An abelian algebra such as $L^\infty(X,\mu)$ is the algebra of bounded measurable functions in a spectral representation, while a nonabelian algebra records simultaneous operator structure beyond a single normal operator. Commutants are the mechanism linking representation theory and spectral multiplicity. If an algebra acts with a large commutant, the representation has internal symmetry; if the centre is small, the representation is closer to being irreducible at the von Neumann algebra level. [/explanation] This operator-theoretic viewpoint explains the early chapters of the course. Weak closure is not a technical decoration: it is the closure operation compatible with spectral projections, normal functionals, and limiting measurement procedures. [explanation: Ergodic Theory And Group Actions] Group measure space constructions attach von Neumann algebras to actions $G\curvearrowright (X,\mathcal E,\mu)$. Ergodicity often corresponds to factoriality, while orbit structure and amenability influence whether the resulting algebra is hyperfinite, rigid, or carries additional classification data. This connection turns dynamical questions into operator-algebraic ones. Instead of studying only points and orbits, one studies the von Neumann algebra generated by functions on the space and unitaries implementing the group action. [/explanation] The group von Neumann algebra $L(G)$ is the special case where the space is a point. Crossed products generalize it and form one of the main routes from measured dynamics into type $II$ and type $III$ examples. [explanation: Quantum Theory And Modular Structure] In finite quantum systems, observables form $M_n(\mathbb C)$ and states are density matrices. Infinite quantum systems lead to representations whose weak closures may be type $II$ or type $III$, and the type determines which operations resemble finite-dimensional quantum mechanics. Modular theory enters when a state is not tracial. The Tomita-Takesaki modular automorphism group records the intrinsic time evolution associated to a faithful normal state, and in type III theory its invariants become part of classification. [/explanation] This is the conceptual reason the course moves from traces to modular theory. Traces flatten left and right multiplication into a symmetric picture; nontracial states produce modular asymmetry, and type III factors make that asymmetry unavoidable. ## Final Summary The type classification of von Neumann algebras is built from projections. Type I algebras retain atomic matrix-like pieces, type II algebras are diffuse but trace-measurable, and type III algebras are purely infinite and governed by modular invariants. The centre decomposes a general algebra into central summands, so factor theory is the local form of the general classification. The course's main tools now fit into one scheme. The bicommutant theorem explains why weakly closed operator algebras are controlled by commutation; projection equivalence supplies the dimension theory; traces distinguish finite and semifinite behavior; modular theory takes over in the purely infinite regime. Later work in the subject refines this scheme through injectivity, hyperfiniteness, crossed products, subfactors, and rigidity phenomena. ## References - J. von Neumann, *On Rings of Operators. Reduction Theory*, Annals of Mathematics, 1938. - F. J. Murray and J. von Neumann, *On Rings of Operators*, Annals of Mathematics, 1936. - M. Takesaki, *Theory of Operator Algebras I*, Springer, 1979. - R. V. Kadison and J. R. Ringrose, *Fundamentals of the Theory of Operator Algebras*, Volumes I-II, Academic Press, 1983-1986. - B. Blackadar, *Operator Algebras: Theory of C*-Algebras and von Neumann Algebras*, Springer, 2006.