39.03.04 · operator-algebras / von-neumann-algebras

Comparison of Projections and the Murray-von Neumann Type Classification

shipped3 tiersLean: none

Anchor (Master): Takesaki *Theory of Operator Algebras I* Ch. V; Murray-von Neumann *On Rings of Operators* I–IV; Dixmier *Von Neumann Algebras* Ch. III; Kadison-Ringrose *Fundamentals II* Ch. 6–8

Intuition Beginner

A projection inside an algebra of operators 39.03.01 picks out a piece of space — it answers a yes-or-no question and keeps the part where the answer is yes. The first thing you want to know about two such pieces is whether they are the same size. In ordinary geometry you compare two regions by sliding one onto the other. Here you do the same thing, but the sliding is done by an operator inside the algebra that carries the first piece exactly onto the second without stretching it. When such a carrier exists, the two projections count as equivalent — the same size in the only sense the algebra can detect.

Once you can compare sizes, you can ask how the sizes are spread out. In familiar settings sizes come in whole numbers: a region holds one point, or two, or seventeen. But operator algebras allow stranger size scales. Some allow a continuous dial of sizes, every value between zero and one, with no smallest nonzero piece at all. Others are so large that every nonzero piece is the same size as the whole, so the scale runs away and swallows itself.

These three behaviors — whole-number sizes, a continuous dial, and a runaway scale — are the heart of the type classification. Every algebra of this kind breaks apart cleanly into pieces of each behavior.

Visual Beginner

A partial isometry is a one-way conveyor belt: it carries one chosen piece of space exactly onto another and ignores everything else. The picture to hold is two shaded regions joined by such a belt, certifying that the two have equal size.

The conveyor belt going both ways is what equality of size means: $v$ takes $p$ onto $q$ , and the reverse belt $v^{*}$ takes $q$ back onto $p$ . The three rulers below show the three ways sizes can be laid out.

Worked example Beginner

Work with three-by-three complex matrices acting on ordinary three-dimensional space. Let $p$ be the projection onto the first coordinate axis and $q$ the projection onto the second coordinate axis:

p = 100000000, q = 000010000 .

Both pick out a one-dimensional piece. To certify they are the same size, find a conveyor belt $v$ carrying the first axis onto the second. Take $v$ to be the matrix sending the first basis vector to the second and killing everything else:

v = 010000000 .

Check the two belt conditions. Multiplying $v^{*} v$ gives back $p$ — the belt loaded at the first axis. Multiplying $v v^{*}$ gives back $q$ — the belt delivers to the second axis. So $p$ and $q$ are equivalent.

What this tells us: in matrices, two projections are the same size exactly when they have the same rank — the same dimension of range. Here both have rank one, and the conveyor belt makes the equal-size verdict concrete. Size is dimension, counted in whole numbers, which is the type I pattern.

Check your understanding Beginner

Formal definition Intermediate+

Let $M \subseteq B (H)$ be a von Neumann algebra 39.03.01 and let $P (M)$ denote its lattice of projections. Two projections $p, q \in P (M)$ are Murray–von Neumann equivalent (relative to $M$ ), written $p \sim q$ , if there exists a partial isometry $v \in M$ with $$ v^* v = p, \qquad v v^* = q. $$ The element $v$ carries the range of $p$ isometrically onto the range of $q$ and annihilates $(ran p)^{⊥}$ . The relation $\sim$ is an equivalence relation on $P (M)$ : reflexivity uses $v = p$ , symmetry uses $v^{*}$ , and transitivity uses the product of two partial isometries. One writes $p ⪯ q$ when $p \sim q^{'}$ for some subprojection $q^{'} \leq q$ ; this is a preorder, and the comparison of projections asks when two projections are $⪯$ -comparable.

A projection $p$ is finite if $p \sim q \leq p$ forces $q = p$ — no proper subprojection is equivalent to the whole. Otherwise $p$ is infinite. A projection $p$ is properly infinite if $p \neq = 0$ and $z p$ is infinite for every central projection $z$ with $z p \neq = 0$ . A nonzero projection $p$ is abelian if the reduced algebra $pM p$ is commutative; equivalently, the only projections under $p$ are central in $pM p$ . The von Neumann algebra $M$ itself is called finite, infinite, properly infinite, or has any of these properties of its unit $1$ .

The centre is $Z (M) = M \cap M^{'}$ 39.03.01. The central support $c (p)$ of a projection $p$ is the smallest central projection $z$ with $z p = p$ ; equivalently $c (p) = ⋁_{u} u p u^{*}$ over unitaries $u \in M$ . The type decomposition is governed by central projections, and the central support is the device that compares projections that live in different central blocks. A projection $p$ is purely infinite in the relevant sense when $M$ contains no nonzero finite central piece below $c (p)$ .

A trace on $M^{+} = {x \in M : x \geq 0}$ is a map $τ : M^{+} \to [0, + \infty]$ that is additive, positively homogeneous, and satisfies the trace property $τ (x^{*} x) = τ (x x^{*})$ for all $x \in M$ . It is faithful if $τ (x) = 0 \Rightarrow x = 0$ , normal if $τ (sup x_{λ}) = sup τ (x_{λ})$ on bounded increasing nets (the order-continuity of 39.03.02), and finite if $τ (1) < \infty$ , in which case it normalises to a state $τ (1) = 1$ . On projections, $p \sim q$ implies $τ (p) = τ (q)$ , so a trace descends to a dimension function on equivalence classes.

Counterexamples to common slips

Equivalence is relative to $M$ , not to $B (H)$ . Two projections can be equivalent in $B (H)$ (same rank) yet inequivalent in a smaller $M$ that contains no partial isometry between them; the partial isometry must live in $M$ . Conversely they can be $M$ -equivalent without being unitarily conjugate by a unitary of $M$ unless their complements are also equivalent.
Finiteness is not the same as finite-dimensional range. In a II $_{1}$ factor the unit $1$ is a finite projection even though $H$ is infinite-dimensional: no proper subprojection of $1$ is equivalent to $1$ , because the trace strictly decreases. The hallmark of infinity is the existence of a proper subprojection of the same size, exactly as the even integers biject with all integers.
The comparison theorem does not say any two projections are comparable outright. It says they become comparable after cutting by a single central projection $z$ : $z p ⪯ z q$ on one side and $(1 - z) q ⪯ (1 - z) p$ on the other. In a factor the centre is the scalars, so $z \in {0, 1}$ and genuine total comparability returns.

Key theorem with proof Intermediate+

Theorem (the comparison theorem). Let $M$ be a von Neumann algebra and $p, q \in P (M)$ . There is a central projection $z \in Z (M)$ such that $$ zp \preceq zq \qquad \text{and} \qquad (1-z)q \preceq (1-z)p. $$ In particular, in a factor any two projections are comparable: either $p ⪯ q$ or $q ⪯ p$ . ^{[Takesaki Ch. V; Murray-von Neumann 1936]}

Proof. The engine is the comparability of projections with equal central support by a maximal-family / halving argument, combined with a central cut that separates the two halves.

Step 1: a partial isometry exists when ranges overlap centrally. First record the basic lemma: if $p, q \in P (M)$ have $c (p) c (q) \neq = 0$ , then there are nonzero subprojections $p_{0} \leq p$ and $q_{0} \leq q$ with $p_{0} \sim q_{0}$ . Indeed, $c (p) c (q) \neq = 0$ means the central supports overlap, so there is $x \in M$ with $q x p \neq = 0$ . Write the polar decomposition $q x p = v ∣ q x p ∣$ with $v \in M$ a partial isometry (the polar parts lie in $M$ because $M$ is weakly closed and $∣ q x p ∣ = (p x^{*} q x p)^{1/2}$ is a norm-limit of polynomials in an element of $M$ ). Then $v^{*} v =$ the support projection of $∣ q x p ∣$ , a nonzero subprojection $p_{0} \leq p$ , and $v v^{*} = q_{0} \leq q$ a nonzero subprojection, with $v$ implementing $p_{0} \sim q_{0}$ .

Step 2: exhaust by a maximal orthogonal family. Consider all families ${(p_{i}, q_{i}, v_{i})}_{i \in I}$ where the $p_{i} \leq p$ are mutually orthogonal, the $q_{i} \leq q$ are mutually orthogonal, and $v_{i}$ implements $p_{i} \sim q_{i}$ . Partially order these families by inclusion; a chain has its union as an upper bound, so Zorn's lemma yields a maximal family. Set $p^{'} = \sum_{i} p_{i} \leq p$ and $q^{'} = \sum_{i} q_{i} \leq q$ . Summing the partial isometries, $v = \sum_{i} v_{i}$ (a strongly convergent sum with orthogonal initial and final spaces) is a partial isometry with $v^{*} v = p^{'}$ and $v v^{*} = q^{'}$ , so $p^{'} \sim q^{'}$ .

Step 3: the leftover pieces have disjoint central support. Let $r = p - p^{'}$ and $s = q - q^{'}$ be the leftovers. By maximality, Step 1 cannot be applied to $r$ and $s$ : there are no nonzero equivalent subprojections of $r$ and $s$ . By the contrapositive of Step 1 this forces $c (r) c (s) = 0$ — the leftovers have orthogonal central supports.

Step 4: the central cut. Put $z = c (s) = 1 -$ (a central projection dominating $r$ off $s$ ). Because $c (r) c (s) = 0$ , we have $z r = 0$ , so $z p = z p^{'} \sim z q^{'} \leq z q$ , giving $z p ⪯ z q$ . On the complement, $(1 - z) s = s - z s$ ; since $z = c (s) \geq s$ , in fact $(1 - z) s = 0$ , so $(1 - z) q = (1 - z) q^{'} \sim (1 - z) p^{'} \leq (1 - z) p$ , giving $(1 - z) q ⪯ (1 - z) p$ . (More carefully, choosing $z$ to be the central projection with $z r = 0$ and $(1 - z) s = 0$ , available because $c (r) c (s) = 0$ lets the two leftovers be separated by a central projection, delivers both inequalities simultaneously.) This is the asserted dichotomy.

For a factor, $Z (M) = C 1$ , so $z \in {0, 1}$ ; one of the two inequalities is vacuous and the other reads $p ⪯ q$ or $q ⪯ p$ . $□$

Bridge. The comparison theorem builds toward the entire type classification, and it appears again in 39.04.01 where the modular theory of a factor refines the comparison of projections into a flow on the projection lattice. The foundational reason the theorem works is exactly the polar decomposition landing inside $M$ : a single operator $q x p \neq = 0$ with overlapping central support manufactures a partial isometry, which is the central insight that turns the abstract question "are these the same size?" into the concrete production of a conveyor belt. This is exactly the bridge from the bicommutant identity $M = M^{''}$ of 39.03.01 — which guarantees weak limits and hence polar parts stay in $M$ — to a quantitative size theory. Putting these together, the centre $Z (M)$ measures precisely the obstruction to total comparability, so a factor (centre the scalars) carries a total order on projection sizes; this generalises the rank function on matrices and is dual to the trace, which reads the order back as a number. The dimension function and the trace, developed next, are the numerical shadow of the order the comparison theorem installs.

Exercises Intermediate+

Exercise 5 (medium, short-answer).

Show that if $τ$ is a faithful finite normal trace on $M$ and $p \sim q$ , then $τ (p) = τ (q)$ ; deduce that the unit of a factor carrying such a trace is finite.

Hint

Use the trace property $τ (v^{*} v) = τ (v v^{*})$ for the implementing partial isometry, then argue strict monotonicity.

Answer

Let $v \in M$ implement $p \sim q$ , so $v^{*} v = p$ , $v v^{*} = q$ . The trace property gives $τ (p) = τ (v^{*} v) = τ (v v^{*}) = τ (q)$ . For finiteness of $1$ : suppose $1 \sim r \leq 1$ . Then $τ (r) = τ (1)$ , so $τ (1 - r) = 0$ ; faithfulness forces $1 - r = 0$ , i.e. $r = 1$ . Hence no proper subprojection of $1$ is equivalent to $1$ , so $1$ is finite. The same argument shows every projection is finite when a faithful finite normal trace is present, so such an $M$ is a finite von Neumann algebra.

Rubric: full credit for the trace-property computation and the faithfulness step forcing $r = 1$ .

Exercise 6 (hard, short-answer).

Prove that $B (H)$ with $dim H = \infty$ is properly infinite by exhibiting a proper subprojection of $1$ equivalent to $1$ .

Hint

Split an orthonormal basis $(e_{n})_{n \geq 1}$ into evens and odds; build a partial isometry from a bijection of index sets.

Answer

Fix an orthonormal basis $(e_{n})_{n \geq 0}$ of $H$ . Let $p$ be the projection onto $\overline{span} {e_{n} : n \geq 1}$ , a proper subprojection of $1$ (it omits $e_{0}$ ). Define $v \in B (H)$ by $v e_{n} = e_{n + 1}$ , the unilateral shift. Then $v^{*} v = 1$ (the shift is an isometry) and $v v^{*} = p$ (the range of $v$ is the span of $e_{1}, e_{2}, \dots$ ). So $1 \sim p < 1$ , proving $1$ is an infinite projection. Because $B (H)$ is a factor, the only central projections are $0$ and $1$ , and $z \cdot 1 = 1$ is infinite for $z = 1$ ; hence $1$ is properly infinite. The same shift construction inside any corner shows every nonzero projection of infinite rank is infinite, so $B (H)$ is the type I $_{\infty}$ factor.

Rubric: full credit for the shift, the identities $v^{*} v = 1$ , $v v^{*} = p < 1$ , and the conclusion of proper infiniteness.

Exercise 7 (hard, short-answer).

Show that a factor $M$ admitting a nonzero abelian projection is of type I, and that the abelian projection has minimal central support behaviour making $M ≅ B (K)$ for some $K$ .

Hint

An abelian projection $e$ has $e M e = C e$ ; use the comparison theorem to tile $1$ by copies of $e$ .

Answer

Let $e \in M$ be abelian, so $e M e$ is commutative; in a factor, $c (e) = 1$ . Abelianness with scalar centre forces $e M e = C e$ : any projection $f \leq e$ lies in $e M e$ and is central there, but $e M e$ being a factor corner with $c (e) = 1$ has centre the scalars, so $f \in {0, e}$ , i.e. $e$ is minimal. By the comparison theorem and $c (e) = 1$ , $1$ is a sum of mutually orthogonal projections each equivalent to $e$ : pick a maximal orthogonal family ${e_{i}}$ of copies of $e$ ; their sum is central (its central support and complement argument forces it to be $1$ in a factor). Indexing this family by a set of cardinality $n$ (possibly infinite) and choosing partial isometries $v_{ij}$ with $v_{ij}^{*} v_{ij} = e_{j}$ , $v_{ij} v_{ij}^{*} = e_{i}$ , the $v_{ij}$ form a system of matrix units, giving a $*$ -isomorphism $M ≅ B (K)$ with $dim K = n$ . So $M$ is type I $_{n}$ (or I $_{\infty}$ ).

Rubric: full credit for minimality of $e$ , tiling $1$ by copies of $e$ , and assembling the matrix units to conclude $M ≅ B (K)$ .

Lean formalization Intermediate+

lean_status: none — Mathlib carries bounded operators, adjoints, and orthogonal projections, and a partial VonNeumannAlgebra predicate, but neither Murray–von Neumann equivalence of projections, the comparison theorem, the finite/infinite/abelian distinctions, nor the type decomposition and the trace on a finite factor is packaged.

The intended formalisation reads schematically:

import Mathlib.Analysis.InnerProductSpace.Adjoint
import Mathlib.Analysis.InnerProductSpace.Projection

variable {H : Type*} [NormedAddCommGroup H] [InnerProductSpace ℂ H] [CompleteSpace H]
variable (M : Set (H →L[ℂ] H)) [IsVonNeumannAlgebra M]

/-- Murray–von Neumann equivalence: a partial isometry in M between p and q. -/
def MvnEquiv (p q : H →L[ℂ] H) : Prop :=
  ∃ v ∈ M, star v * v = p ∧ v * star v = q

/-- The comparison theorem: any two projections become comparable after a
central cut z ∈ Z(M). -/
theorem comparison
    (p q : H →L[ℂ] H) (hp : IsProjection M p) (hq : IsProjection M q) :
    ∃ z ∈ center M, IsCentralProjection z ∧
      Subordinate (MvnEquiv M) (z * p) (z * q) ∧
      Subordinate (MvnEquiv M) ((1 - z) * q) ((1 - z) * p) :=
  sorry  -- maximal-family exhaustion + polar decomposition in M

Advanced results Master

Comparison of projections installs an order on sizes; the trace turns the order into a number, and the central decomposition sorts the algebra into the three behaviours.

The type decomposition. Every von Neumann algebra $M$ admits a unique decomposition by central projections $1 = z_{I} + z_{II} + z_{III}$ , with $z_{I} M$ of type I (it has an abelian projection of central support $1$ ), $z_{II} M$ of type II (no nonzero abelian projection, but a finite projection of central support $1$ ), and $z_{III} M$ of type III (no nonzero finite projection). The type II part splits further as $z_{II_{1}} + z_{II_{\infty}}$ according to whether $1$ is finite or properly infinite, and the type I part stratifies as $⨁_{n} z_{I_{n}} M$ with $z_{I_{n}} M$ a homogeneous algebra $≅ B (K_{n}) \overset{ˉ}{\otimes} A_{n}$ for abelian $A_{n}$ and $dim K_{n} = n \in {1, 2, \dots, \infty}$ . A factor ( $Z (M) = C 1$ ) is purely of one type: I $_{n}$ , I $_{\infty}$ , II $_{1}$ , II $_{\infty}$ , or III.

Factors of each type. Type I $_{n}$ is $M_{n} (C)$ ; type I $_{\infty}$ is $B (H)$ . Type II $_{1}$ is the prototype of a finite factor of infinite linear dimension — the group von Neumann algebra $L (G)$ of an i.c.c. group 39.03.01, or the hyperfinite $R$ below. Type II $_{\infty}$ is $II_{1} \overset{ˉ}{\otimes} B (H)$ , equivalently $R \overset{ˉ}{\otimes} B (ℓ^{2})$ , carrying a semifinite trace with range $[0, \infty]$ . Type III factors carry no trace at all; the Powers factors $R_{λ}$ ( $0 < λ < 1$ ), built from infinite tensor products of $2 \times 2$ matrix states with eigenvalue ratio $λ$ , are mutually non-isomorphic type III factors, and Connes' invariants $S (M)$ and $T (M)$ refine III into the continuum III $_{λ}$ , $0 \leq λ \leq 1$ 39.04.01.

The dimension function and the trace on a finite factor. On a finite factor $M$ there is a unique faithful normal trace $τ$ with $τ (1) = 1$ , and the dimension function $d (p) = τ (p)$ on projections realises the comparison order numerically: $p ⪯ q ⟺ d (p) \leq d (q)$ , and $p \sim q ⟺ d (p) = d (q)$ . Its range is ${0, \frac{1}{n}, \dots, 1}$ for type I $_{n}$ and the whole interval $[0, 1]$ for type II $_{1}$ . This continuous range — a projection of every size between $0$ and $1$ , hence no minimal projection — is the defining novelty Murray and von Neumann discovered: "continuous dimension". Uniqueness of $τ$ follows because any two normal traces on a factor are proportional (the comparison theorem makes the trace determined by its values on equivalence classes, which the order pins down up to scale).

Construction and uniqueness of the hyperfinite II $_{1}$ factor. Let $R$ be the weak closure of the increasing union $⋃_{n} M_{2^{n}} (C)$ of dyadic matrix algebras, embedded $x \mapsto x \otimes 1_{2}$ , in the GNS representation of the unique normalised trace. Then $R$ is a II $_{1}$ factor, hyperfinite (a weak limit of finite-dimensional subalgebras) and approximately finite-dimensional. Murray and von Neumann proved $R$ is the unique hyperfinite II $_{1}$ factor up to isomorphism; Connes' deep theorem later identified hyperfiniteness with injectivity, amenability, and several other conditions, making $R$ the canonical "smallest" II $_{1}$ factor into which every separable amenable one embeds.

The fundamental group. For a II $_{1}$ factor $M$ with trace $τ$ , the amplifications $M_{t} = M \overset{ˉ}{\otimes} B (ℓ^{2})$ cut by a projection of trace $t$ define isomorphism types $M^{t}$ for $t > 0$ , and the fundamental group $F (M) = {t > 0 : M^{t} ≅ M}$ is a subgroup of $R_{\times}^{+}$ measuring the self-similarity of $M$ under rescaling the trace. For the hyperfinite factor $F (R) = R_{\times}^{+}$ (it is fully self-similar); Connes showed property-(T) factors $L (G)$ have countable $F$ , and Popa later produced II $_{1}$ factors with $F (M) = {1}$ and with prescribed fundamental groups, opening the deformation/rigidity theory.

Existence of type III. Type III factors are not pathological exceptions: they are generic. The Powers factors $R_{λ}$ exist by the infinite-tensor-product construction, the group-measure-space algebra $L^{\infty} (X) ⋊ G$ of a non-measure-preserving ergodic action is type III, and the local algebras of relativistic quantum field theory are type III $_{1}$ 39.04.01. The Tomita–Takesaki modular automorphism group, built from a faithful normal state via 39.03.02, is the tool that detects and classifies them, since a type III factor admits no trace to anchor the comparison theory to a number.

Synthesis. Comparison of projections is the foundational reason von Neumann algebras carry a structure theory at all: the comparison theorem installs a total order on projection sizes inside a factor, and this is exactly what the type classification reads off — discrete order (type I), continuous order with a trace (type II), or an order with no finite values (type III). The dimension function generalises matrix rank, and putting these together the trace $τ$ is dual to the order, recovering $p ⪯ q ⟺ τ (p) \leq τ (q)$ as the numerical face of the comparison theorem. The central insight is that the centre $Z (M)$ is the precise obstruction both to comparability and to being a single factor, so the central decomposition into types and the direct-integral decomposition into factors 39.03.01 are one and the same cut; this is exactly the bridge from the bicommutant identity of 39.03.01 and the predual duality of 39.03.02 to a complete-invariant program. The continuous dimension of II $_{1}$ and the tracelessness of III generalise, respectively, the finite-rank and the runaway scales, and they hand to modular theory 39.04.01 the type III factors that no trace can reach — the deformation/rigidity invariants such as the fundamental group then animate the bare order with a moduli theory.

Full proof set Master

Proposition (equivalence is an equivalence relation). On $P (M)$ the relation $\sim$ is reflexive, symmetric, and transitive. Proof. Reflexivity: $v = p$ has $v^{*} v = p = v v^{*}$ . Symmetry: if $v^{*} v = p$ , $v v^{*} = q$ , then $w = v^{*}$ has $w^{*} w = v v^{*} = q$ , $w w^{*} = v^{*} v = p$ , so $q \sim p$ . Transitivity: if $p \sim q$ via $v$ and $q \sim r$ via $w$ (so $w^{*} w = q = v v^{*}$ ), set $u = w v$ . Then $u^{*} u = v^{*} w^{*} w v = v^{*} q v = v^{*} (v v^{*}) v = (v^{*} v) (v^{*} v) = p$ since $v^{*} v = p$ is a projection, and $u u^{*} = w v v^{*} w^{*} = w q w^{*} = w (w^{*} w) w^{*} = q^{'}$ ; computing, $w q w^{*} = w (w^{*} w) w^{*} = (w w^{*}) (w w^{*}) = w w^{*} = r$ . Hence $u$ implements $p \sim r$ . $□$

Proposition (additivity of equivalence over orthogonal families). If ${p_{i}}$ and ${q_{i}}$ are orthogonal families with $p_{i} \sim q_{i}$ via $v_{i} \in M$ , then $\sum_{i} p_{i} \sim \sum_{i} q_{i}$ . Proof. The partial isometries $v_{i}$ have mutually orthogonal initial spaces $p_{i}$ and mutually orthogonal final spaces $q_{i}$ . The sum $v = \sum_{i} v_{i}$ converges strongly and lies in $M$ (a strong limit of finite partial sums, each in $M$ , and $M$ is SOT-closed 39.03.01). Orthogonality gives $v^{*} v = \sum_{i} v_{i}^{*} v_{i} = \sum_{i} p_{i}$ and $v v^{*} = \sum_{i} v_{i} v_{i}^{*} = \sum_{i} q_{i}$ , the cross terms vanishing because $v_{j}^{*} v_{i} = 0$ for $i \neq = j$ (ranges of $v_{i}$ orthogonal) and similarly for $v_{i} v_{j}^{*}$ . $□$

Proposition (a factor with a faithful normal finite trace is finite, and the trace is unique up to scale). Let $M$ be a factor with a faithful normal trace $τ$ , $τ (1) = 1$ . Then $M$ is finite and any normal trace $τ^{'}$ on $M$ equals $τ^{'} (1) τ$ . Proof. Finiteness is Exercise 5. For uniqueness, by the comparison theorem any two projections $p, q$ are comparable, so the function $d (p) = τ (p)$ is a dimension function: $p \sim q \Rightarrow d (p) = d (q)$ and $p ⊥ q \Rightarrow d (p + q) = d (p) + d (q)$ . A normal trace $τ^{'}$ likewise yields $d^{'} (p) = τ^{'} (p)$ with the same additivity and equivalence-invariance. On a factor the comparison order is total, and any two normal dimension functions agreeing on the total order are proportional: for dyadic-rational sizes one halves $1$ repeatedly (possible since $τ$ takes all values in $[0, 1]$ for II $_{1}$ , or all of ${k / n}$ for I $_{n}$ ), forcing $d^{'} = d^{'} (1) d$ there, and normality (order-continuity) extends the equality to all projections; linearity and density of the span of projections extend it to all of $M$ . Hence $τ^{'} = τ^{'} (1) τ$ . $□$

Proposition (the type decomposition is unique). There exist unique central projections $z_{I}, z_{II}, z_{III}$ summing to $1$ such that $z_{I} M$ is type I, $z_{II} M$ type II, $z_{III} M$ type III. Proof. Let $z_{I}$ be the central support of the supremum of all abelian projections (this supremum's central support is central). Then $z_{I} M$ has an abelian projection of full central support (type I), and $(1 - z_{I}) M$ has none. Within $(1 - z_{I}) M$ , let $z_{II}$ be the central support of the supremum of all finite projections; $(1 - z_{I}) z_{II} M$ then has a finite projection of full central support and no abelian one (type II), while the remainder $z_{III} = 1 - z_{I} - z_{II}$ supports no nonzero finite projection (type III). Uniqueness: the defining properties (existence/non-existence of abelian and finite projections of full central support) are intrinsic to the reduced algebra, so any central projection splitting $M$ into these types must agree with $z_{I}, z_{II}, z_{III}$ on each block; a central projection determined by an intrinsic property of its reduced algebra is unique. The further split $z_{II} = z_{II_{1}} + z_{II_{\infty}}$ uses finiteness versus proper infiniteness of the unit, equally intrinsic. $□$

Proposition (II $_{1}$ has continuous dimension: every value in $[0, 1]$ is attained). Let $M$ be a II $_{1}$ factor with trace $τ$ and $t \in [0, 1]$ . There is a projection $p \in M$ with $τ (p) = t$ . Proof. The trace is normal and faithful, and $M$ has no minimal projection (a minimal projection would be abelian, making $M$ type I). Given any projection $q$ with $τ (q) > 0$ , non-minimality yields $0 \neq = q^{'} < q$ , and the comparison theorem inside the factor $q M q$ lets one halve: there is $q^{''} \leq q$ with $τ (q^{''}) = \frac{1}{2} τ (q)$ (compare $q^{''}$ and $q - q^{''}$ ; equality of the two halves is achievable because the order is total and the trace continuous). Iterating produces projections of every dyadic-rational trace in $[0, τ (1)] = [0, 1]$ . For arbitrary $t$ , take an increasing sequence of dyadic projections $p_{n}$ with $τ (p_{n}) ↑ t$ and orthogonal increments; their supremum $p = ⋁ p_{n}$ satisfies $τ (p) = lim τ (p_{n}) = t$ by normality. $□$

Proposition (the hyperfinite II $_{1}$ factor is a factor). The weak closure $R = \overline{⋃_{n} M_{2^{n}} (C)}$ in the GNS representation of the unique tracial state is a II $_{1}$ factor. Proof. The union $A = ⋃_{n} M_{2^{n}}$ is a unital $*$ -algebra with a faithful tracial state $τ$ (the consistent normalised matrix traces); its GNS representation $π_{τ}$ has a cyclic and separating vector $Ω$ , and $R = π_{τ} (A)^{''}$ is a von Neumann algebra carrying the normal extension of $τ$ 39.03.02. The trace is faithful and finite, so $R$ is finite. Factoriality: an element $x \in Z (R)$ commutes with every $M_{2^{n}}$ , hence with the whole union $A$ ; since $A$ is weakly dense in $R$ , $x$ commutes with $R$ , and a tracial-state computation shows $x$ commutes with all matrix units in every $M_{2^{n}}$ , forcing $x$ to be scalar on each finite block and, by trace continuity, scalar overall: $Z (R) = C 1$ . Type: $R$ is infinite-dimensional (it contains $M_{2^{n}}$ for all $n$ ) and finite with a faithful normal trace, so it is neither I $_{n}$ nor I $_{\infty}$ nor properly infinite; it is II $_{1}$ . $□$

Connections Master

Von Neumann algebras and the bicommutant theorem 39.03.01 — the comparison theorem rests on partial isometries from polar decompositions remaining inside $M$ , which is guaranteed by SOT-closedness $M = M^{''}$ ; the central decomposition into types is the projection-lattice refinement of the direct-integral decomposition of $M$ over its centre introduced there, and the II $_{1}$ factor $L (G)$ of an i.c.c. group is the running example carried over.
The predual, normal states, and the σ-weak topology 39.03.02 — the trace on a finite factor is a faithful normal state, hence an element of the predual $M_{*}$ , and normality (order-continuity) is exactly what makes the dimension function attain suprema and the continuous-dimension construction close; the GNS representation of this tracial state builds the hyperfinite factor $R$ .
Tomita–Takesaki modular theory 39.04.01 — type III factors carry no trace, so the comparison theory cannot be numericised by a dimension function; the modular automorphism group of a faithful normal state is the substitute, and Connes' invariants $S (M)$ , $T (M)$ split type III into the continuum III $_{λ}$ , completing the classification the type decomposition begins here.
C-algebras: axioms, spectrum, functional calculus 39.01.01* — finite-dimensional C*-algebras are exactly the type I $_{n}$ direct sums $⨁_{k} M_{n_{k}} (C)$ , the simplest finite von Neumann algebras, and the matrix-unit assembly that proves a type I factor is $B (K)$ is the same device that classifies finite-dimensional C*-algebras.

Historical & philosophical context Master

The comparison theory and the type classification are due to Murray and von Neumann in their series On Rings of Operators (1936–1943) ^{[Murray-von Neumann 1936]}. The first paper introduced equivalence of projections by partial isometries and the relative dimension function, and discovered that a factor need not have its projection sizes quantised: alongside the type I $_{n}$ algebras with dimension range ${0, 1, \dots, n}$ they constructed a factor whose dimension takes every value in $[0, 1]$ , the type II $_{1}$ case, and named the phenomenon "continuous dimension". The construction of the first II $_{1}$ factor used the regular representation of a group with infinite conjugacy classes, and the third paper (1943) proved that the approximately finite-dimensional II $_{1}$ factor is unique — the result that singles out the hyperfinite factor $R$ .

The existence of type III, where no trace exists and the comparison theory cannot be anchored to a number, was established in the same series and long resisted finer analysis. Powers (1967) constructed a one-parameter family of mutually non-isomorphic type III factors by infinite tensor products, showing type III is richly populated. Connes (1973) classified type III factors into III $_{λ}$ , $0 \leq λ \leq 1$ , via the modular spectrum and the invariants $S (M)$ and $T (M)$ extracted from Tomita–Takesaki theory, and proved the uniqueness of the injective (hyperfinite) factor of each type, for which he received the Fields Medal in 1982. The fundamental group, introduced by Murray and von Neumann as a measure of self-similarity, became computable and controllable only with Popa's deformation/rigidity theory in the 2000s, which produced II $_{1}$ factors with fundamental group ${1}$ and with arbitrary prescribed fundamental groups.

Bibliography Master

Takesaki, M., Theory of Operator Algebras I, Springer, 1979. Ch. V.
Murray, F. J. and von Neumann, J., "On rings of operators", Annals of Mathematics 37 (1936), 116–229.
Murray, F. J. and von Neumann, J., "On rings of operators. IV", Annals of Mathematics 44 (1943), 716–808.
Dixmier, J., Von Neumann Algebras, North-Holland, 1981. Ch. III.
Powers, R. T., "Representations of uniformly hyperfinite algebras and their associated von Neumann rings", Annals of Mathematics 86 (1967), 138–171.
Connes, A., "Une classification des facteurs de type III", Annales scientifiques de l'École Normale Supérieure 6 (1973), 133–252.
Kadison, R. V. and Ringrose, J. R., Fundamentals of the Theory of Operator Algebras II: Advanced Theory, Academic Press, 1986. Ch. 6–8.

Operator-algebras spine, von Neumann-algebra structure unit. Comparison of projections by partial isometries and the comparison theorem as the order on sizes that the Murray–von Neumann type classification (I/II/III) reads off; builds on the bicommutant theory (39.03.01) and the predual/normal-state machinery (39.03.02), and hands the traceless type III factors to modular theory (39.04.01).

Prerequisites

39.03.01
39.03.02

Tier anchors

beginner: Strogatz-style 'two projections are the same size when one can be slid onto the other by a partial isometry', with the type classification as a coarse measure of how the sizes spread out — like whole numbers, a continuous dial, or a runaway scale that swallows itself
intermediate: Takesaki *Theory of Operator Algebras I* Ch. V §1–2; Conway *A Course in Operator Theory* Ch. X; Murphy *C*-Algebras and Operator Theory* Ch. 4
master: Takesaki *Theory of Operator Algebras I* Ch. V; Murray-von Neumann *On Rings of Operators* I–IV; Dixmier *Von Neumann Algebras* Ch. III; Kadison-Ringrose *Fundamentals II* Ch. 6–8

References

Takesaki, M. — Theory of Operator Algebras I · Ch. V, comparison of projections, the comparison theorem, the type decomposition I/II/III, traces on finite von Neumann algebras
Murray, F. J. and von Neumann, J. — On Rings of Operators · Ann. of Math. 37 (1936), 116–229 (equivalence and comparison of projections, the dimension function); Trans. AMS 41 (1937); Ann. of Math. 44 (1943) (the uniqueness of the hyperfinite II_1 factor)
Dixmier, J. — Von Neumann Algebras · Ch. III, the relative dimension theory, the trace, and the type decomposition

Estimated time

beginner: 20m
intermediate: 55m
master: 90m