Introduction
Connes’ bicentralizer conjecture and the weak Dixmier property
In the theory of von Neumann algebras, as in many areas of mathematics, averaging and averagable objects are a central theme of research. Amenability is the property that describes the possibility of averaging and is a central property in geometric and ergodic group theory. A discrete group Γ \Gamma Γ is amenable if and only if there exists a Γ \Gamma Γ -invariant mean on ℓ ∞ ( Γ ) \ell^\infty(\Gamma) ℓ ∞ ( Γ ) [42 ] .
For von Neumann algebras, injectivity is the appropriate notion of amenability . A von Neumann algebra M ⊂ B ( H ) M \subset B(\mathcal{H}) M ⊂ B ( H ) is injective if and only if there exists a conditional expectation from B ( H ) B(\mathcal{H}) B ( H ) onto M M M . In his celebrated work, Connes [7 ] proved that, for von Neumann algebras with separable predual, injectivity is equivalent to being hyperfinite , which means that the algebra can be approximated by finite-dimensional subalgebras. For hyperfinite M M M , we can use these approximations to average over the compact unitary groups of finite-dimensional subalgebras. Passing to weak limits, we can average elements of B ( H ) B(\mathcal{H}) B ( H ) into M ′ M' M ′ . That is, M ⊂ B ( H ) M \subset B(\mathcal{H}) M ⊂ B ( H ) has the weak Dixmier property, which is an averaging property of inclusions of von Neumann algebras in the following sense: a W ∗ W^* W ∗ -inclusion M ⊂ A M \subset\mathcal{A} M ⊂ A , i.e., an inclusion of von Neumann algebras, has the weak Dixmier property if every x ∈ A x \in\mathcal{A} x ∈ A can be averaged by unitaries of M M M into the relative commutant M ′ ∩ A M' \cap\mathcal{A} M ′ ∩ A , that is,
conv ‾ w { u x u ∗ ∣ u ∈ U ( M ) } ∩ ( M ′ ∩ A ) ≠ ∅ ( x ∈ A ) . \overline{\operatorname{conv}}^{w}\{uxu^{*} \mid u \in\mathcal{U}(M)\} \cap(M' \cap\mathcal{A}) \ne\emptyset\qquad(x \in\mathcal{A}). conv w { ux u ∗ ∣ u ∈ U ( M )} ∩ ( M ′ ∩ A ) = ∅ ( x ∈ A ) .
This connection between finite-dimensional approximation and averaging is a tool for the classification of amenable von Neumann algebras.
The classification of von Neumann algebras was a central goal of the work of Murray–von Neumann [MvN36]. Von Neumann algebras with separable predual decompose as direct integrals of factors , which are von Neumann algebras M M M with trivial center, i.e., Z ( M ) = M ′ ∩ M = C 1 Z(M)=M' \cap M=\mathbb{C}1 Z ( M ) = M ′ ∩ M = C 1 . Murray–von Neumann divided factors into types I, II1 _1 1 , II∞ _\infty ∞ , and III. A type I factor is isomorphic to B ( H ) B(\mathcal{H}) B ( H ) for some Hilbert space H \mathcal{H} H . A type II1 _1 1 factor admits a normal faithful finite trace, while a type II∞ _\infty ∞ factor admits a normal faithful semifinite trace τ \tau τ with τ ( 1 ) = ∞ \tau(1)=\infty τ ( 1 ) = ∞ . Type III factors admit no nonzero normal semifinite trace. Murray–von Neumann [35 ] proved that, up to isomorphism, there is a unique hyperfinite II1 _1 1 factor.
Connes [6 ] refined this classification by dividing type III into types IIIλ _\lambda λ for 0 ≤ λ ≤ 1 0 \leq\lambda\leq1 0 ≤ λ ≤ 1 . He subsequently classified the amenable type II1 _1 1 , type II∞ _\infty ∞ , and type IIIλ _\lambda λ factors with separable predual for 0 ≤ λ < 1 0 \leq\lambda< 1 0 ≤ λ < 1 [7 ] . For the remaining type III1 _1 1 case, Connes [8 ] introduced the bicentralizer conjecture and showed that a positive solution for amenable factors would imply the uniqueness of the amenable type III1 _1 1 factor.
Let M M M be a countably decomposable von Neumann algebra and let ϕ \phi ϕ be a normal faithful state on M M M . The bicentralizer B ( M , ϕ ) B(M,\phi) B ( M , ϕ ) consists of the x ∈ M x \in M x ∈ M such that [ x , a n ] → 0 [x,a_n] \to0 [ x , a n ] → 0 strongly for every bounded sequence ( a n ) ⊂ M (a_n) \subset M ( a n ) ⊂ M satisfying ∥ a n ϕ − ϕ a n ∥ → 0 \|a_n\phi-\phi a_n\| \to0 ∥ a n ϕ − ϕ a n ∥ → 0 , see Subsection 2.3. Connes’ bicentralizer conjecture states that B ( M , ϕ ) = C 1 B(M,\phi)=\mathbb{C}1 B ( M , ϕ ) = C 1 for every type III1 _1 1 factor M M M and every normal faithful state ϕ \phi ϕ on M M M .
The finite trace τ \tau τ on a type II1 _1 1 factor M M M makes the analogous bicentralizer statement immediate: B ( M , τ ) = Z ( M ) = C 1 B(M,\tau)=Z(M)=\mathbb{C}1 B ( M , τ ) = Z ( M ) = C 1 . More generally, much of the power of averaging in finite von Neumann algebras comes from the trace: it provides an invariant Hilbert-space geometry in which convexity produces fixed points. Type III factors have no such trace. For type IIIλ _\lambda λ factors with separable predual, 0 < λ < 1 0<\lambda<1 0 < λ < 1 , by the periodic modular structure we have a discrete decomposition as a crossed product of a type II∞ _\infty ∞ factor by Z \mathbb{Z} Z [6 ] . This reduces many structural questions to the semifinite setting. Such a periodic reduction is not available in general for type III1 _1 1 factors. One must therefore find methods that work without either a trace on the original algebra or such a discrete reduction. This is one reason why the type III1 _1 1 case often presents the greatest difficulty.
The same difficulty appears in Popa’s problem on the weak Dixmier property (Corollary E ) and Kadison’s problem on MASA (Theorem G), discussed below. For inclusions M ⊂ A M \subset\mathcal{A} M ⊂ A with expectation and finite M M M , the weak Dixmier property follows from classical averaging arguments. The type II1 _1 1 case of Kadison’s problem was proved by Popa [38 ] in 1981. The type III1 _1 1 cases of these problems were resolved only much later, in 2019 and 2026, respectively.
Haagerup [20 Theorem 2.3] solved the bicentralizer problem for amenable factors, thereby proving the uniqueness of the amenable type III1 _1 1 factor and completing the classification of amenable factors with separable predual. Haagerup [20 Theorem 3.1] proved that a type III1 _1 1 factor M M M with separable predual has trivial bicentralizer if and only if M ⊂ c ( M ) M \subset c(M) M ⊂ c ( M ) has the weak Dixmier property, where c ( M ) = M ⋊ σ ϕ R c(M)=M\rtimes_{\sigma^\phi}\mathbb{R} c ( M ) = M ⋊ σ ϕ R is the continuous core of M M M . Here σ ϕ \sigma^\phi σ ϕ is the modular automorphism group associated with ϕ \phi ϕ , see Subsection 2.2 for background on modular theory and continuous cores. For amenable M M M , by hyperfiniteness we can average each element of c ( M ) c(M) c ( M ) by unitaries of M M M into M ′ ∩ c ( M ) M' \cap c(M) M ′ ∩ c ( M ) in the weak topology, so M ⊂ c ( M ) M \subset c(M) M ⊂ c ( M ) has the weak Dixmier property and hence M M M has trivial bicentralizer.
Houdayer–Marrakchi [HM26, Theorem C] recently solved Connes’ bicentralizer conjecture for arbitrary type III1 _1 1 factors. Their proof does not follow Haagerup’s approach. It relies on the bicentralizer flow [AHHM20], on the ergodicity theorems for this flow proved by Marrakchi [30 Theorem D] , [33 Theorem C] , and on a resonance phenomenon, discovered in [HM26], for the joint action of the modular flow and the dual bicentralizer flow. In particular, their proof does not use Haagerup’s theorem. It is therefore also a new proof that amenable type I I I 1 \mathrm{III}_1 III 1 factors have trivial bicentralizer, and hence, by [8 ] , a new proof of the uniqueness of the amenable type I I I 1 \mathrm{III}_1 III 1 factor.
In this paper, we follow Haagerup’s approach to give a new proof of Connes’ bicentralizer conjecture, independent of [HM26]. We prove the weak Dixmier property for M ⊂ c ( M ) M \subset c(M) M ⊂ c ( M ) directly, and more generally for M ⊂ c ( P ) M \subset c(P) M ⊂ c ( P ) whenever M ⊂ P M \subset P M ⊂ P admits a normal faithful conditional expectation, and then apply Haagerup’s characterization. Our proof treats all factors uniformly, with only minor adjustments between the finite and properly infinite cases. In particular, the previously most difficult type I I I 1 \mathrm{III}_1 III 1 case is treated by the same argument as the other properly infinite cases, without using any recent advances in modular and bicentralizer theory.
Theorem A. Let M ⊂ P M \subset P M ⊂ P be an inclusion of countably decomposable von Neumann algebras with a normal faithful conditional expectation. Then M ⊂ c ( P ) M \subset c(P) M ⊂ c ( P ) has the weak Dixmier property. In particular, M ⊂ c ( M ) M \subset c(M) M ⊂ c ( M ) has the weak Dixmier property.
By the result of Haagerup [20 Theorem 3.1] , Theorem A implies Connes’ bicentralizer conjecture.
Corollary B (Connes’ bicentralizer conjecture [HM26]). Let M M M be a type I I I 1 \mathrm{III}_1 III 1 factor with separable predual. Then B ( M , ϕ ) = C 1 B(M,\phi)=\mathbb{C}1 B ( M , ϕ ) = C 1 for every normal faithful state ϕ \phi ϕ on M M M .
By [32 Theorem D] , Theorem A also implies Marrakchi’s relative bicentralizer conjecture [32 Conjecture C] , which generalizes Connes’ bicentralizer conjecture. We refer to Subsection 2.3 for the definition of the relative bicentralizer and the statement of the relative bicentralizer conjecture.
Corollary C (The relative bicentralizer conjecture). Let M ⊂ P M \subset P M ⊂ P be an inclusion of countably decomposable von Neumann algebras with a normal faithful conditional expectation. Then M ⊂ P M \subset P M ⊂ P satisfies the relative bicentralizer conjecture.
Our proof of Theorem A is based on Poisson boundary techniques. To this end, we extend the Poisson boundary theory of Das–Peterson [9 ] from finite von Neumann algebras to general von Neumann algebras with separable preduals. Subsection 1.2 introduces this theory, starting from the averaging interpretation of Poisson boundaries of groups, and explains how we use it to prove the weak Dixmier property. Subsection 1.3 discusses another application of this theory, to Kadison’s problem on MASA. Subsection 1.4 explains the proof strategy for Theorem A.
The weak Dixmier property via Poisson boundaries
We first recall Poisson boundaries for groups, which are also objects defined through averaging. Let Γ \Gamma Γ be a countable discrete group, and let μ ∈ Prob ( Γ ) \mu\in\operatorname{Prob}(\Gamma) μ ∈ Prob ( Γ ) be a generating probability measure, i.e., ⋃ n ≥ 1 ( supp μ ) n = Γ \bigcup_{n\geq1}(\operatorname{supp}\mu)^n=\Gamma ⋃ n ≥ 1 ( supp μ ) n = Γ . Define the Markov operator P μ P_\mu P μ on ℓ ∞ ( Γ ) \ell^\infty(\Gamma) ℓ ∞ ( Γ ) by
P μ ( f ) ( g ) = ∑ s ∈ Γ μ ( s ) f ( g s ) ( g ∈ Γ ) . P_\mu(f)(g)=\sum_{s\in\Gamma}\mu(s)f(gs) \qquad(g\in\Gamma). P μ ( f ) ( g ) = s ∈ Γ ∑ μ ( s ) f ( g s ) ( g ∈ Γ ) .
Hence P μ P_\mu P μ can be viewed as a μ \mu μ -weighted averaging operator. Let Har ( Γ , μ ) = { f ∈ ℓ ∞ ( Γ ) ∣ P μ ( f ) = f } \operatorname{Har}(\Gamma,\mu)=\{f\in\ell^\infty(\Gamma)\mid P_\mu(f)=f\} Har ( Γ , μ ) = { f ∈ ℓ ∞ ( Γ ) ∣ P μ ( f ) = f } be the space of bounded μ \mu μ -harmonic functions.
The Poisson boundary of ( Γ , μ ) (\Gamma,\mu) ( Γ , μ ) is the essentially unique nonsingular Γ \Gamma Γ -probability space ( B , ν B ) (B,\nu_B) ( B , ν B ) such that the Poisson transform
P ν B : L ∞ ( B , ν B ) ⟶ Har ( Γ , μ ) , P ν B ( F ) ( g ) = ∫ B F ( g b ) d ν B ( b ) P_{\nu_B}:L^\infty(B,\nu_B)\longrightarrow\operatorname{Har}(\Gamma,\mu),\qquad P_{\nu_B}(F)(g)=\int_B F(gb)\,d\nu_B(b) P ν B : L ∞ ( B , ν B ) ⟶ Har ( Γ , μ ) , P ν B ( F ) ( g ) = ∫ B F ( g b ) d ν B ( b )
is an isometric isomorphism [12 ] .
For example, let Γ \Gamma Γ be a non-elementary hyperbolic group with a finite generating set S S S , and let μ \mu μ be a generating probability measure supported on S S S . Then its Poisson boundary can be identified with the Gromov boundary ( ∂ ∞ Γ , ν ∞ ) (\partial_\infty\Gamma,\nu_\infty) ( ∂ ∞ Γ , ν ∞ ) of the Cayley graph with respect to S S S , where ν ∞ \nu_\infty ν ∞ is the hitting measure of the μ \mu μ -random walk [28 Theorems 7.6 and 7.7] .
Another example comes from lattices in semisimple Lie groups. Furstenberg [13 ] proved that, for every irreducible lattice Γ < G \Gamma< G Γ < G in a connected semisimple real Lie group with finite center and no compact factors, there exists a fully supported probability measure μ \mu μ on Γ \Gamma Γ such that the ( Γ , μ ) (\Gamma,\mu) ( Γ , μ ) -Poisson boundary is ( G / P , ν P ) (G/P,\nu_P) ( G / P , ν P ) , where P < G P < G P < G is a minimal parabolic subgroup. A major application of Poisson boundary theory is the proof of Margulis’ superrigidity theorem for irreducible lattices in higher rank semisimple Lie groups [29 Chapter VII] .
We now turn to the corresponding construction for von Neumann algebras. Izumi [24 , 25 ] developed a noncommutative Poisson boundary construction for normal ucp maps on von Neumann algebras, motivated by the study of discrete quantum groups. Building on Izumi’s construction, Das–Peterson [9 ] initiated Poisson boundary theory for finite von Neumann algebras, which was subsequently studied in [Zho24b, Zho24a, Zho24c, Zho25]. We extend this theory to general von Neumann algebras M M M with separable preduals. In this general setting, as analogues of generating probability measures on groups, our main objects are normal, regular, strongly generating (nrsg) hyperstates on B ( L 2 ( M ) ) B(L^{2}(M)) B ( L 2 ( M )) (Definition Definition 3.8 ). More concretely (see Example 3.9 ), given any normal faithful state ϕ \phi ϕ on M M M with canonical cyclic vector ξ ϕ ∈ L 2 ( M ) \xi_\phi\in L^{2}(M) ξ ϕ ∈ L 2 ( M ) , we can always construct a normal state φ \varphi φ on B ( L 2 ( M ) ) B(L^{2}(M)) B ( L 2 ( M )) with φ ∣ M = ϕ \varphi|_{M}=\phi φ ∣ M = ϕ and sequences ( z k ) , ( y k ) ⊂ M (z_k),(y_k)\subset M ( z k ) , ( y k ) ⊂ M such that
C [ ( z k ) ] ‾ w = C [ ( y k ) ] ‾ w = M , ∑ k z k ∗ z k = ∑ k y k y k ∗ = 1. \overline{\mathcal{C}[(z_k)]}^{\,w}=\overline{\mathcal{C}[(y_k)]}^{\,w}=M,\qquad\sum_k z_k^{*}z_k=\sum_k y_ky_k^{*}=1. C [( z k )] w = C [( y k )] w = M , k ∑ z k ∗ z k = k ∑ y k y k ∗ = 1.
The corresponding normal inner ucp maps
P φ ( T ) = ∑ k ( J z k ∗ J ) T ( J z k J ) , Q φ ( T ) = ∑ k y k T y k ∗ (1) P_\varphi(T)=\sum_k(Jz_k^{*}J)T(Jz_kJ),\qquad Q_\varphi(T)=\sum_k y_kTy_k^{*} \tag*{(1)} P φ ( T ) = k ∑ ( J z k ∗ J ) T ( J z k J ) , Q φ ( T ) = k ∑ y k T y k ∗ (1)
satisfy
φ ( T ) = ⟨ P φ ( T ) ξ ϕ , ξ ϕ ⟩ = ⟨ Q φ ( T ) ξ ϕ , ξ ϕ ⟩ ( T ∈ B ( L 2 ( M ) ) ) . (2) \varphi(T)=\langle P_\varphi(T)\xi_\phi,\xi_\phi\rangle=\langle Q_\varphi(T)\xi_\phi,\xi_\phi\rangle\qquad(T\in B(L^{2}(M))). \tag*{(2)} φ ( T ) = ⟨ P φ ( T ) ξ ϕ , ξ ϕ ⟩ = ⟨ Q φ ( T ) ξ ϕ , ξ ϕ ⟩ ( T ∈ B ( L 2 ( M ))) . (2)
Here J J J is the modular conjugation on L 2 ( M ) L^{2}(M) L 2 ( M ) , and C [ ( z k ) ] \mathcal{C}[(z_k)] C [( z k )] and C [ ( y k ) ] \mathcal{C}[(y_k)] C [( y k )] are the unital subalgebras (rather than the unital ∗ * ∗ -subalgebras) generated by ( z k ) (z_k) ( z k ) and ( y k ) (y_k) ( y k ) , respectively.
Let Har ( P φ ) = { T ∈ B ( L 2 ( M ) ) ∣ P φ ( T ) = T } \operatorname{Har}(P_\varphi)=\{T\in B(L^{2}(M))\mid P_\varphi(T)=T\} Har ( P φ ) = { T ∈ B ( L 2 ( M )) ∣ P φ ( T ) = T } . We define the Poisson boundary B φ \mathcal{B}_\varphi B φ to be the von Neumann algebra isomorphic to Har ( P φ ) \operatorname{Har}(P_\varphi) Har ( P φ ) as operator systems (Definition Definition 3.3 ). Its existence and uniqueness are proved in Proposition Proposition 3.4 . Under this identification, M M M is a von Neumann subalgebra of B φ \mathcal{B}_\varphi B φ .
For example, let Γ \Gamma Γ be a countable discrete group, let μ ∈ Prob ( Γ ) \mu\in\operatorname{Prob}(\Gamma) μ ∈ Prob ( Γ ) be a generating probability measure, and let ( B , ν B ) (B,\nu_B) ( B , ν B ) be its Poisson boundary. For M = L ( Γ ) M=L(\Gamma) M = L ( Γ ) , consider the hyperstate φ ∈ B ( ℓ 2 ( Γ ) ) ∗ \varphi\in B(\ell^{2}(\Gamma))^{*} φ ∈ B ( ℓ 2 ( Γ ) ) ∗ defined by φ ( T ) = ∑ g ∈ Γ μ ( g ) ⟨ T δ g , δ g ⟩ \varphi(T)=\sum_{g\in\Gamma}\mu(g)\langle T\delta_g,\delta_g\rangle φ ( T ) = ∑ g ∈ Γ μ ( g ) ⟨ T δ g , δ g ⟩ . Izumi [25 ] proved that M ⊂ B φ M\subset\mathcal{B}_\varphi M ⊂ B φ is naturally isomorphic to the inclusion L ( Γ ) ⊂ L ∞ ( B , ν B ) ⋊ Γ L(\Gamma)\subset L^\infty(B,\nu_B)\rtimes\Gamma L ( Γ ) ⊂ L ∞ ( B , ν B ) ⋊ Γ .
The following result is the starting point of our approach to the weak Dixmier property. We deduce this property from the existence of a suitable map into B φ \mathcal{B}_\varphi B φ .
Theorem D (Theorem 4.2 ). Let M M M be a von Neumann algebra with separable predual, and let φ \varphi φ be an nrsg hyperstate on B ( L 2 ( M ) ) B(L^{2}(M)) B ( L 2 ( M )) with Poisson boundary B φ \mathcal{B}_\varphi B φ . Assume that one of the following conditions holds:
(i) M M M is properly infinite, i.e., M = M ∞ ⊕ M I I ∞ ⊕ M I I I M=M_{\infty}\oplus M_{\mathrm{II}_{\infty}}\oplus M_{\mathrm{III}} M = M ∞ ⊕ M II ∞ ⊕ M III ;
(ii) M M M is finite with a normal faithful trace τ \tau τ , and φ = ∫ U ( M ) ⟨ ⋅ ξ τ , ξ τ ⟩ d μ ( u ) \varphi=\int_{U(M)}\langle\cdot\,\xi_\tau,\xi_\tau\rangle\,d\mu(u) φ = ∫ U ( M ) ⟨ ⋅ ξ τ , ξ τ ⟩ d μ ( u ) for some μ ∈ Prob ( U ( M ) ) \mu\in\operatorname{Prob}(U(M)) μ ∈ Prob ( U ( M )) .
Then every W ∗ W^{*} W ∗ -inclusion M ⊂ A M\subset A M ⊂ A equipped with a normal faithful ucp map Φ : A → B φ \Phi:A\to\mathcal{B}_\varphi Φ : A → B φ satisfying Φ ∣ M = id M \Phi|_M=\operatorname{id}_M Φ ∣ M = id M has the weak Dixmier property.
As a direct corollary, we give a new proof of the weak Dixmier property for W ∗ W^{*} W ∗ -inclusions with a normal faithful conditional expectation, a problem posed by Popa [39 Remark 1.5] and solved by Marrakchi [31 ] .
Corollary E ([39 , 31 ] , Corollary 4.3). Let M ⊂ A M \subset A M ⊂ A be an inclusion of countably decomposable von Neumann algebras with a normal faithful conditional expectation. Then M ⊂ A M \subset A M ⊂ A has the weak Dixmier property.
We also prove a semifinite version of Theorem D, replacing the normal faithful ucp map into B φ \mathcal{B}_{\varphi} B φ by a normal faithful semifinite M M M -bimodular positive map T : A + → B ^ φ + T:\mathcal{A}_{+} \to\widehat{\mathcal{B}}_{\varphi+} T : A + → B φ + , with the averaging conclusion for positive elements x x x such that T ( x ) T(x) T ( x ) is bounded (Theorem 4.7 ). In particular, we have a new proof of the weak Dixmier result of Isono [Iso25] for inclusions M ⊂ A M \subset A M ⊂ A of countably decomposable von Neumann algebras with a normal faithful semifinite operator-valued weight (Corollary 4.8 ).
The proof of Theorem D is rather simple. We explain the key case where M M M is properly infinite with separable predual, and there is a normal faithful conditional expectation E : A → M E:A \to M E : A → M . Fix a normal faithful state ϕ \phi ϕ on M M M and take an nrsg hyperstate φ \varphi φ extending ϕ \phi ϕ as above. Using the coefficients ( y k ) (y_k) ( y k ) of Q φ Q_{\varphi} Q φ as in (1) , we define the normal M M M -inner ucp map
Q : A → A , Q ( a ) = ∑ k y k a y k ∗ ( a ∈ A ) , Q:\mathcal{A} \to\mathcal{A}, \qquad Q(a)=\sum_k y_k a y_k^* \qquad(a \in\mathcal{A}), Q : A → A , Q ( a ) = k ∑ y k a y k ∗ ( a ∈ A ) ,
and let Har ( Q , A ) = { a ∈ A ∣ Q ( a ) = a } \operatorname{Har}(Q,\mathcal{A})=\{a\in\mathcal{A}\mid Q(a)=a\} Har ( Q , A ) = { a ∈ A ∣ Q ( a ) = a } . We first prove that Har ( Q , A ) = M ′ ∩ A \operatorname{Har}(Q,\mathcal{A})=M'\cap\mathcal{A} Har ( Q , A ) = M ′ ∩ A .
Define ψ A = ϕ ∘ E \psi_{\mathcal{A}}=\phi\circ E ψ A = ϕ ∘ E . For a ∈ A a\in\mathcal{A} a ∈ A with Q ( a ) = a Q(a)=a Q ( a ) = a , by Kadison’s inequality we have
Q ( a a ∗ ) − a a ∗ = Q ( a a ∗ ) − Q ( a ) Q ( a ) ∗ ≥ 0. Q(aa^*)-aa^*=Q(aa^*)-Q(a)Q(a)^* \ge0. Q ( a a ∗ ) − a a ∗ = Q ( a a ∗ ) − Q ( a ) Q ( a ) ∗ ≥ 0.
By (2) , ϕ ∘ Q ∣ M = ϕ \phi\circ Q|_M=\phi ϕ ∘ Q ∣ M = ϕ . Since E E E is M M M -bimodular, we also have
ψ A ∘ Q = ϕ ∘ E ∘ Q = ϕ ∘ Q ∘ E = ϕ ∘ E = ψ A . (3) \psi_{\mathcal{A}}\circ Q=\phi\circ E\circ Q=\phi\circ Q\circ E=\phi\circ E=\psi_{\mathcal{A}}. \tag*{(3)} ψ A ∘ Q = ϕ ∘ E ∘ Q = ϕ ∘ Q ∘ E = ϕ ∘ E = ψ A . (3)
Hence ψ A ( Q ( a a ∗ ) − a a ∗ ) = 0 \psi_{\mathcal{A}}(Q(aa^*)-aa^*)=0 ψ A ( Q ( a a ∗ ) − a a ∗ ) = 0 , and by faithfulness we have Q ( a a ∗ ) = a a ∗ Q(aa^*)=aa^* Q ( a a ∗ ) = a a ∗ . Since Q ( a ∗ ) = a ∗ Q(a^*)=a^* Q ( a ∗ ) = a ∗ , we have
∑ k ( y k a − a y k ) ( y k a − a y k ) ∗ = ∑ k ( y k a a ∗ y k ∗ − y k a y k ∗ a ∗ − a y k a ∗ y k ∗ + a y k y k ∗ a ∗ ) = Q ( a a ∗ ) − Q ( a ) a ∗ − a Q ( a ∗ ) + a a ∗ = 0. (4) \begin{aligned}
\sum_k (y_k a-a y_k)(y_k a-a y_k)^* \\
&=\sum_k (y_k aa^*y_k^*-y_kay_k^*a^*-ay_k a^*y_k^*+ay_ky_k^*a^*) \\
&=Q(aa^*)-Q(a)a^*-aQ(a^*)+aa^*=0. \tag*{(4)}
\end{aligned} k ∑ ( y k a − a y k ) ( y k a − a y k ) ∗ = k ∑ ( y k a a ∗ y k ∗ − y k a y k ∗ a ∗ − a y k a ∗ y k ∗ + a y k y k ∗ a ∗ ) = Q ( a a ∗ ) − Q ( a ) a ∗ − a Q ( a ∗ ) + a a ∗ = 0. (4)
Hence y k a = a y k y_k a=a y_k y k a = a y k for every k k k . Since C ∗ [ ( y k ) ] C^*[(y_k)] C ∗ [( y k )] is weakly dense in M M M , we have a ∈ M ′ ∩ A a\in M'\cap\mathcal{A} a ∈ M ′ ∩ A , proving Har ( Q , A ) = M ′ ∩ A \operatorname{Har}(Q,\mathcal{A})=M'\cap\mathcal{A} Har ( Q , A ) = M ′ ∩ A .
For every a ∈ A a\in\mathcal{A} a ∈ A , any weak cluster point of the Cesàro averages C n ( a ) = ( 1 / n ) ∑ j = 0 n − 1 Q j ( a ) C_n(a)=(1/n)\sum_{j=0}^{n-1}Q^j(a) C n ( a ) = ( 1/ n ) ∑ j = 0 n − 1 Q j ( a ) lies in Har ( Q , A ) = M ′ ∩ A \operatorname{Har}(Q,\mathcal{A})=M'\cap\mathcal{A} Har ( Q , A ) = M ′ ∩ A . If y k = ∥ y k ∥ u k y_k=\lVert y_k\rVert u_k y k = ∥ y k ∥ u k with u k ∈ U ( M ) u_k\in U(M) u k ∈ U ( M ) , these averages would lie in the weakly closed convex hull of the unitary orbit of a a a , which would prove the weak Dixmier property for M ⊂ A M\subset\mathcal{A} M ⊂ A directly. In general, the coefficients ( y k ) (y_k) ( y k ) need not have this form. Nevertheless, the same argument shows that M ⊂ A M\subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property defined below.
Definition 1.1. Let M ⊂ A M\subset\mathcal{A} M ⊂ A be a W ∗ W^* W ∗ -inclusion. A ucp map P : A → A P:\mathcal{A}\to\mathcal{A} P : A → A is called M M M -inner if
P ( a ) = ∑ k z k a z k ∗ , ∑ k z k z k ∗ = 1 , P(a)=\sum_k z_kaz_k^*, \qquad\sum_k z_kz_k^*=1, P ( a ) = k ∑ z k a z k ∗ , k ∑ z k z k ∗ = 1 , for a sequence ( z k ) ⊂ M (z_k)\subset M ( z k ) ⊂ M . We say that M ⊂ A M\subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property if, for every a ∈ A a\in\mathcal{A} a ∈ A ,
{ P ( a ) ∣ P is an M -inner ucp map } ‾ w ∩ ( M ′ ∩ A ) ≠ ∅ . \overline{\{P(a)\mid P\text{ is an }M\text{-inner ucp map}\}}^{\,w}\cap(M'\cap\mathcal{A})\ne\varnothing. { P ( a ) ∣ P is an M -inner ucp map } w ∩ ( M ′ ∩ A ) = ∅ .
Indeed, each Cesàro averaging map C n C_n C n is M M M -inner, so M ⊂ A M\subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property. We prove that the two properties are equivalent when M M M is properly infinite, which completes the proof in the case considered above. The proof uses proper infiniteness to pass from inner ucp maps to unitary averages, an idea similar to that used by Haagerup [19 Section 2] .
Theorem F (Theorem 4.1). Let M ⊂ A M \subset\mathcal{A} M ⊂ A be a W ∗ W^* W ∗ -inclusion such that M M M is properly infinite. Then M ⊂ A M \subset\mathcal{A} M ⊂ A has the weak Dixmier property if and only if it has the inner UCP-Dixmier property.
We briefly explain the proof of Theorem F. We only need to prove the if direction. By proper infiniteness, we can identify M ⊂ A M \subset\mathcal{A} M ⊂ A with M ⊗ ‾ B ( ℓ 2 ( N ) ) ⊂ A ⊗ ‾ B ( ℓ 2 ( N ) ) M\overline{\otimes}B(\ell^2(\mathbb{N})) \subset\mathcal{A}\overline{\otimes}B(\ell^2(\mathbb{N})) M ⊗ B ( ℓ 2 ( N )) ⊂ A ⊗ B ( ℓ 2 ( N )) . We only need to prove the weak Dixmier property for the latter inclusion. Since B ( ℓ 2 ( N ) ) B(\ell^2(\mathbb{N})) B ( ℓ 2 ( N )) is hyperfinite, every x ∈ A ⊗ ‾ B ( ℓ 2 ( N ) ) x \in\mathcal{A}\overline{\otimes}B(\ell^2(\mathbb{N})) x ∈ A ⊗ B ( ℓ 2 ( N )) can be averaged by unitaries of 1 ⊗ B ( ℓ 2 ( N ) ) 1 \otimes B(\ell^2(\mathbb{N})) 1 ⊗ B ( ℓ 2 ( N )) to some a ⊗ 1 a \otimes1 a ⊗ 1 with a ∈ A a \in\mathcal{A} a ∈ A . By the inner UCP-Dixmier property of M ⊂ A M \subset\mathcal{A} M ⊂ A , we only need to prove that, for every M M M -inner ucp map P : A → A \mathcal{P}:\mathcal{A}\to\mathcal{A} P : A → A ,
P ( a ) ⊗ 1 ∈ conv ‾ w { u ( a ⊗ 1 ) u ∗ ∣ u ∈ U ( M ⊗ ‾ B ( ℓ 2 ( N ) ) ) } . \mathcal{P}(a)\otimes1\in\overline{\operatorname{conv}}^{\,w}\{u(a\otimes1)u^* \mid u\in\mathcal{U}(M\overline{\otimes}B(\ell^2(\mathbb{N})))\}. P ( a ) ⊗ 1 ∈ conv w { u ( a ⊗ 1 ) u ∗ ∣ u ∈ U ( M ⊗ B ( ℓ 2 ( N )))} .
Indeed, we can then average x x x into ( M ′ ∩ A ) ⊗ ‾ C 1 (M' \cap\mathcal{A})\overline{\otimes}\mathbb{C}1 ( M ′ ∩ A ) ⊗ C 1 , the relative commutant of the amplified inclusion.
Write P ( a ) = ∑ k ≥ 1 z k a z k ∗ \mathcal{P}(a)=\sum_{k\geq1}z_kaz_k^* P ( a ) = ∑ k ≥ 1 z k a z k ∗ with ( z k ) ⊂ M (z_k)\subset M ( z k ) ⊂ M and ∑ k z k z k ∗ = 1 \sum_k z_kz_k^*=1 ∑ k z k z k ∗ = 1 . For n ≥ 1 n\geq1 n ≥ 1 , let Z n = ( 0 n × n , z 1 I n , z 2 I n , … ) Z_n=(0_{n\times n},z_1I_n,z_2I_n,\ldots) Z n = ( 0 n × n , z 1 I n , z 2 I n , … ) . Since M M M is properly infinite, we can complete this block-row to a unitary U n ∈ M ⊗ ‾ B ( ℓ 2 ( N ) ) U_n\in M\overline{\otimes}B(\ell^2(\mathbb{N})) U n ∈ M ⊗ B ( ℓ 2 ( N )) with first block-row Z n Z_n Z n , so that
U n = ( Z n ∗ ) and U n ( a ⊗ 1 ) U n ∗ = ( P ( a ) I n ∗ ∗ ∗ ) . U_n=\begin{pmatrix} Z_n \\ * \end{pmatrix}
\qquad\text{and}\qquad
U_n(a\otimes1)U_n^*=
\begin{pmatrix}
\mathcal{P}(a)I_n & *\\
* & *
\end{pmatrix}. U n = ( Z n ∗ ) and U n ( a ⊗ 1 ) U n ∗ = ( P ( a ) I n ∗ ∗ ∗ ) .
Then we have U n ( a ⊗ 1 ) U n ∗ → P ( a ) ⊗ 1 U_n(a\otimes1)U_n^*\to\mathcal{P}(a)\otimes1 U n ( a ⊗ 1 ) U n ∗ → P ( a ) ⊗ 1 weakly as n → ∞ n\to\infty n → ∞ , which proves the desired inclusion.
To summarize, the Poisson boundary theory provides inner ucp maps with the properties needed to prove the inner UCP-Dixmier property in Definition 1.1; when M M M is properly infinite, we then apply Theorem F to prove the weak Dixmier property.
Kadison’s problem on MASA
The same Poisson boundary tools also give a direct approach to Kadison’s problem on maximal abelian subalgebras (MASA). In 1967, Kadison [27 ] asked whether, for every inclusion of factors M ⊂ P M\subset P M ⊂ P with M ′ ∩ P = C 1 M'\cap P=\mathbb{C}1 M ′ ∩ P = C 1 , there is a maximal abelian subalgebra A ⊂ M A\subset M A ⊂ M which is also maximal abelian in P P P , that is, A ′ ∩ P = A A'\cap P=A A ′ ∩ P = A . In 1981, Popa [38 ] answered this question affirmatively when M M M is of type II and there is a normal faithful conditional expectation from P P P onto M M M , by constructing the desired maximal abelian subalgebra as the limit of an increasing sequence of finite-dimensional abelian subalgebras. Ge–Popa [GP98, Corollary 5.1] showed that the answer can be negative without such a conditional expectation, even when M M M is a type II1 _1 1 factor and P P P is a type II∞ _\infty ∞ factor, and asked whether the existence of a normal faithful conditional expectation suffices for factors of arbitrary type, see the discussion after [GP98, Corollary 5.1] and [40 Problem 7.6] . Marrakchi [32 Theorem A] solved this problem when M M M is not of type III1 _1 1 , or when M M M is a type III1 _1 1 factor with trivial bicentralizer. Marrakchi [33 Theorem A(2)] completed the solution for arbitrary type III1 _1 1 factors, by reducing to the existence of a common irreducible hyperfinite subalgebra [33 Theorem B] , which relies on the ergodicity of the relative bicentralizer flow [33 Theorem C] .
We give a new proof of the solution to Kadison’s problem, different from the proofs of Marrakchi [32 , 33 ] . As in our proof of Theorem A, the most difficult type III1 _1 1 case is treated by the same argument as the other continuous properly infinite cases, without using any recent advances in modular and bicentralizer theory. Moreover, although both proofs use the same Poisson boundary framework, this proof is independent of our proof of the bicentralizer conjecture (Theorem A and Corollaries B and C).
Theorem G ([38 , 32 , 33 ] ). Let M ⊂ P M\subset P M ⊂ P be an inclusion of von Neumann algebras with separable preduals. Assume that M ′ ∩ P = Z ( M ) M'\cap P=\mathcal{Z}(M) M ′ ∩ P = Z ( M ) and that there is a normal faithful conditional expectation E : P → M E:P\to M E : P → M . Then there is an abelian von Neumann subalgebra A ⊂ M A\subset M A ⊂ M such that A ′ ∩ P = A A'\cap P=A A ′ ∩ P = A .
In the key case where M M M is continuous and properly infinite, we construct finite-dimensional subalgebras F n ⊂ M F_n\subset M F n ⊂ M such that ∥ θ ∣ F n ′ ∩ P ∥ → 0 \|\theta|_{F_n'\cap P}\|\to0 ∥ θ ∣ F n ′ ∩ P ∥ → 0 as n → ∞ n\to\infty n → ∞ for every θ ∈ P ∗ \theta\in P_* θ ∈ P ∗ with θ ∣ M ′ ∩ P = 0 \theta|_{M'\cap P}=0 θ ∣ M ′ ∩ P = 0 . Once we have this, the rest follows the well-established route of Popa [38 ] , also used in [32 , 33 ] . We construct the F n F_n F n directly using Har ( Q , P ) = M ′ ∩ P \operatorname{Har}(Q,P)=M'\cap P Har ( Q , P ) = M ′ ∩ P for the inner ucp map φ \varphi φ associated with an rnsg hyperstate, as in Subsection 1.2, and a construction similar to that in the proof of Theorem F; while Marrakchi [Mar26, Theorem B] first constructs a hyperfinite subalgebra N ⊂ M N \subset M N ⊂ M with N ′ ∩ P = M ′ ∩ P N' \cap P = M' \cap P N ′ ∩ P = M ′ ∩ P and then takes ( F n ) (F_n) ( F n ) to be an increasing sequence of finite-dimensional subalgebras such that N = ⋁ n F n N = \bigvee_n F_n N = ⋁ n F n . See also [38 Theorem 3.2 and Remark 3.5(2)] for the construction of such a hyperfinite subalgebra N ⊂ M N \subset M N ⊂ M when M M M is semifinite. In Marrakchi’s proof, the ergodicity of the relative bicentralizer flow [33 Theorem C] was used in the type I I I 1 \mathrm{III}_1 III 1 case. In our proof, the corresponding ergodicity input is the identity Har ( Q , P ) = M ′ ∩ P \operatorname{Har}(Q,P) = M' \cap P Har ( Q , P ) = M ′ ∩ P .
With our Poisson boundary tools, we can in fact prove the following stronger result for properly infinite M M M , replacing the conditional expectation by a normal faithful ucp map into a Poisson boundary. It generalizes the hyperfinite-subalgebra existence results of Popa and Marrakchi in this case and also gives a new proof of them.
Theorem H (Theorem 6.10 ). Let M ⊂ P M \subset P M ⊂ P be an inclusion of von Neumann algebras with separable preduals, and assume that M M M is properly infinite. Let φ \varphi φ be an nrsg M M M -hyperstate. If there is a normal faithful ucp map Φ : P → B φ \Phi: P \to B_\varphi Φ : P → B φ with Φ ∣ M = id M \Phi|_M = \operatorname{id}_M Φ ∣ M = id M , then there is a hyperfinite von Neumann subalgebra N ⊂ M N \subset M N ⊂ M such that N ′ ∩ P = M ′ ∩ P N' \cap P = M' \cap P N ′ ∩ P = M ′ ∩ P .
Note that Theorem H also generalizes Theorem D(i) for inclusions with separable preduals. Indeed, since N N N is hyperfinite, every element of P P P can be averaged by unitaries in U ( N ) ⊂ U ( M ) \mathcal{U}(N) \subset\mathcal{U}(M) U ( N ) ⊂ U ( M ) into N ′ ∩ P = M ′ ∩ P N' \cap P = M' \cap P N ′ ∩ P = M ′ ∩ P . Hence M ⊂ P M \subset P M ⊂ P has the weak Dixmier property.
Proof strategy for Theorem A
Let M ⊂ P M \subset P M ⊂ P be a W ∗ W^* W ∗ -inclusion with a normal faithful conditional expectation E : P → M E : P \to M E : P → M as in Theorem A.
We again focus on the key case where M M M is properly infinite with separable predual. Fix a normal faithful state ϕ \phi ϕ on M M M , and let ρ = ϕ ∘ E \rho= \phi\circ E ρ = ϕ ∘ E . In Subsections 1.2 and 1.3 an arbitrary nrsg hyperstate suffices. Here we take the specific nrsg ( M , ϕ ) (M,\phi) ( M , ϕ ) -hyperstate φ \varphi φ from Example 3.9, with the symmetry φ ( J a ∗ J ) = φ ( a ) \varphi(Ja^*J) = \varphi(a) φ ( J a ∗ J ) = φ ( a ) (a ∈ B ( L 2 ( M ) ) a \in B(L^2(M)) a ∈ B ( L 2 ( M )) ). This is the natural construction used to prove the existence of nrsg hyperstates. It also gives two further properties: the coefficient family ( y k ) (y_k) ( y k ) of Q φ Q_\varphi Q φ is permuted by y k ↦ σ i / 2 φ ( y k ∗ ) y_k \mapsto\sigma_{i/2}^\varphi(y_k^*) y k ↦ σ i /2 φ ( y k ∗ ) , and the operator norms of the y k y_k y k satisfy ∑ k ∥ y k ∥ ∞ 2 < ∞ \sum_k \lVert y_k\rVert_\infty^2 < \infty ∑ k ∥ y k ∥ ∞ 2 < ∞ , see the discussion before Example 3.9. These are exactly the two properties needed for Proposition 1.2 below.
Let A = c ρ ( P ) \mathcal{A} = c_\rho(P) A = c ρ ( P ) , and define Q ( x ) = ∑ k y k x y k ∗ Q(x) = \sum_k y_k x y_k^* Q ( x ) = ∑ k y k x y k ∗ (x ∈ A x \in\mathcal{A} x ∈ A ) using the coefficients ( y k ) (y_k) ( y k ) of Q φ Q_\varphi Q φ . As in Subsection 1.2, to prove the weak Dixmier property for M ⊂ A M \subset\mathcal{A} M ⊂ A , we only need to prove that Har ( Q , A ) = M ′ ∩ A \operatorname{Har}(Q,\mathcal{A}) = M' \cap\mathcal{A} Har ( Q , A ) = M ′ ∩ A .
Let T : A + → P ^ + T : \mathcal{A}_+ \to\widehat{P}_+ T : A + → P + be the canonical normal faithful semifinite operator-valued weight of the continuous core. Define ψ A = ρ ∘ T \psi_{\mathcal{A}} = \rho\circ T ψ A = ρ ∘ T , the dual weight of ρ \rho ρ , see Subsection 2.2.
By (4) , to prove Har ( Q , A ) = M ′ ∩ A \operatorname{Har}(Q,\mathcal{A}) = M' \cap\mathcal{A} Har ( Q , A ) = M ′ ∩ A , we only need to prove that Q ( a a ∗ ) = a a ∗ Q(aa^*) = aa^* Q ( a a ∗ ) = a a ∗ for every a ∈ Har ( Q , A ) a \in\operatorname{Har}(Q,\mathcal{A}) a ∈ Har ( Q , A ) . Fix such an a a a , and let d : = Q ( a a ∗ ) − a a ∗ ≥ 0 d := Q(aa^*) - aa^* \geq0 d := Q ( a a ∗ ) − a a ∗ ≥ 0 .
As in (3) , we have ψ A ∘ Q = ψ A \psi_{\mathcal{A}} \circ Q = \psi_{\mathcal{A}} ψ A ∘ Q = ψ A . Intuitively, we would have ψ A ( d ) = ψ A ( Q ( a a ∗ ) ) − ψ A ( a a ∗ ) = 0 \psi_{\mathcal{A}}(d) = \psi_{\mathcal{A}}(Q(aa^*)) - \psi_{\mathcal{A}}(aa^*) = 0 ψ A ( d ) = ψ A ( Q ( a a ∗ )) − ψ A ( a a ∗ ) = 0 , but this calculation is justified only when ψ A ( a a ∗ ) < ∞ \psi_{\mathcal{A}}(aa^*) < \infty ψ A ( a a ∗ ) < ∞ . Instead, Lemma 2.2 shows that, for every 0 ≠ d ′ ∈ A + 0 \ne d' \in\mathcal{A}_+ 0 = d ′ ∈ A + , there exists x ′ ∈ A + x' \in\mathcal{A}_+ x ′ ∈ A + such that 0 < x ′ ≤ d ′ 0 < x' \leq d' 0 < x ′ ≤ d ′ and 0 < ψ A ( x ′ ) < ∞ 0 < \psi_{\mathcal{A}}(x') < \infty 0 < ψ A ( x ′ ) < ∞ . Hence we only need to prove that every x ∈ A + x \in\mathcal{A}_+ x ∈ A + satisfying x ≤ d x \leq d x ≤ d and ψ A ( x ) < ∞ \psi_{\mathcal{A}}(x) < \infty ψ A ( x ) < ∞ must satisfy x = 0 x = 0 x = 0 . Since x ≤ d = Q ( a a ∗ ) − a a ∗ x \leq d = Q(aa^*) - aa^* x ≤ d = Q ( a a ∗ ) − a a ∗ and ∑ k = 0 n − 1 Q k ( d ) = Q n ( a a ∗ ) − a a ∗ \sum_{k=0}^{n-1} Q^k(d) = Q^n(aa^*) - aa^* ∑ k = 0 n − 1 Q k ( d ) = Q n ( a a ∗ ) − a a ∗ , we have ∑ k = 0 n − 1 Q k ( x ) ≤ C 1 \sum_{k=0}^{n-1} Q^k(x) \leq C1 ∑ k = 0 n − 1 Q k ( x ) ≤ C 1 for every n ≥ 1 n \geq1 n ≥ 1 , with C = ∥ a ∥ 2 C = \lVert a\rVert^2 C = ∥ a ∥ 2 .
All we need to prove is the following proposition.
Proposition 1.2 (Lemmas 5.4 and 5.5). Let A \mathcal{A} A , ψ A \psi_{\mathcal{A}} ψ A and Q Q Q be as above, and let x ∈ A + x \in\mathcal{A}_+ x ∈ A + satisfy ψ A ( x ) < ∞ \psi_{\mathcal{A}}(x) < \infty ψ A ( x ) < ∞ . Assume that there is a constant 0 ≤ C < ∞ 0 \leq C < \infty 0 ≤ C < ∞ such that
∑ k = 0 n − 1 Q k ( x ) ≤ C 1 ( n ≥ 1 ) . \sum_{k=0}^{n-1} Q^k(x) \leq C1 \qquad(n \geq1). k = 0 ∑ n − 1 Q k ( x ) ≤ C 1 ( n ≥ 1 ) . Then x = 0 x = 0 x = 0 .
The estimates in the proof of Proposition 1.2 use a noncommutative analogue of the classical cutoff argument for recurrence of symmetric Markov operators [11 Section 1.6, Theorem 1.6.3] . We explain this argument through the commutative model in Proposition 5.1 . The symmetry of the nrsg hyperstate φ \varphi φ is essential here. Without it, the vanishing conclusion already fails in the commutative model, see Remark 5.2.
Acknowledgments. The author thanks Cyril Houdayer, Amine Marrakchi, Soham Chakraborty and Stuart White for many helpful discussions and encouragement. Proposition 3.2 benefited from a short note written by Soham Chakraborty and Se-Jin (Sam) Kim. Theorem 4.1 benefited from a discussion with Amine Marrakchi during the author’s visit to ENS Lyon in October 2025. Part of this work (§ 3 and § 4 ) was completed while the author was a PhD student at ENS Paris.
Independent and concurrent work. Our solution to the relative bicentralizer conjecture (Theorem A and Corollary C) was obtained in parallel with and independently of the concurrent work [36 ] , released on 6 October 2026. The two proofs use completely different approaches. OpenAI’s proof builds on the work of Houdayer–Marrakchi [HM26], while ours uses the Poisson boundary techniques developed here. See the subsection on AI-assisted research methodology below for the timeline.
AI-assisted research methodology. It is well known that AI can help solve a given problem by identifying techniques whose relevance had previously gone unnoticed. Our solution to the relative bicentralizer conjecture (Theorem A), however, illustrates the reverse process: starting from a given body of techniques developed by the human author, AI helped identify a solvable problem whose connection with these techniques the author had not previously noticed.
The underlying Poisson boundary theory in § 3 was developed entirely by the human author between February 2025 and June 2025, benefiting from helpful discussions with Cyril Houdayer and Soham Chakraborty.
The new proof for Popa’s problem on the weak Dixmier property (Corollary E ) was also purely human work. During the author’s visit to ENS Lyon in October 2025, helpful discussions with Amine Marrakchi led to the observation that the framework in § 3 could be used to simplify the proof of this result. Between October 2025 and May 2026, the author further generalized this result to inclusions with maps into Poisson boundaries (Theorem D) and to the semifinite setting of Isono’s result (Corollary 4.8 ), and prepared a manuscript containing the material of the present § 3 and § 4 . A $20 version of Gemini 3.1 Pro served only as a writing aid during April 2026 (for parts that remain in the present text). During subsequent checking of this manuscript, GPT-6 Astra helped identify and repair minor gaps.
In September 2026, the author began using GPT-6 Astra to investigate whether every type II1 _1 1 factor admits a type III1 _1 1 Poisson boundary, a question that had interested the author since the early stages of his doctoral studies. This investigation did not resolve the original question. The author then asked the model to consider the weaker conclusion that every type II1 _1 1 factor admits an irreducible embedding into the hyperfinite type III1 _1 1 factor. This led to a construction using Poisson boundaries and crossed products. To extend this construction to arbitrary continuous factors, the author gave the model his earlier Poisson boundary manuscript on general von Neumann algebras. The model then helped develop the corresponding embedding arguments. To keep this paper concise and focused, these irreducible embedding results are not included in the present version. We refer interested readers to a preliminary version [Zho26] of this paper.
Having recognized Poisson boundaries as a very powerful tool, the author then asked GPT-6 Astra on 26 September 2026 which other open problems might be solved using the existing Poisson boundary theory. The model suggested three problems that were still open at the time, before OpenAI’s public release of 722 math manuscripts on 6 October 2026 [37 ] :
the relative bicentralizer conjecture;
the free group factor problem;
the vanishing cohomology conjecture of Kadison–Ringrose.
The author chose the relative bicentralizer conjecture for further study. The main tools needed for the proof, including the Poisson boundary construction, symmetric nrsg hyperstates, and Theorem F, had already been developed in the earlier Poisson boundary manuscript. GPT-6 Astra helped connect these tools to the conjecture and work out their application. With this assistance, the author obtained a preliminary proof on 27 September 2026.
The human author identified Kadison’s problem (Theorem G) as another application of the Poisson boundary tools: Following the successful treatment of Popa’s problem on the weak Dixmier property and the relative bicentralizer conjecture, the author recognized the parallels between these two problems: both are formulated for general von Neumann algebras, with classical type I I 1 \mathrm{II}_1 II 1 cases but extremely difficult type I I I 1 \mathrm{III}_1 III 1 cases. The author saw the same contrast in Kadison’s problem on MASA and therefore proposed applying these tools to it. The proof uses the generating inner ucp maps constructed in the author’s earlier Poisson boundary manuscript and the unitary dilation argument used in Theorem F. With the assistance of GPT-6 Astra in applying these tools to Kadison’s problem, the author obtained a substantially simplified proof on 9 October 2026.
Lean was also used to check a formalization of a short note consisting of the preliminaries in Subsection 2.2, Example 3.9 , Theorem 4.1, and the whole of § 6 . These components form the main line of the proof of Theorem A. Results from the published literature were taken as explicit assumptions.
The author is responsible for the mathematical claims, citations, and final presentation.
Preliminaries
The standard form of von Neumann algebras
Let M M M be a countably decomposable von Neumann algebra. Following [15 ] , the standard form of M M M is a quadruple ( M , H , J , H + ) (M,\mathcal{H},J,\mathcal{H}_{+}) ( M , H , J , H + ) such that H \mathcal{H} H is a Hilbert space with M ⊂ B ( H ) M \subset B(\mathcal{H}) M ⊂ B ( H ) ; J : H → H J:\mathcal{H}\to\mathcal{H} J : H → H is an anti-unitary with J 2 = 1 J^{2}=1 J 2 = 1 ; H + ⊂ H \mathcal{H}_{+}\subset\mathcal{H} H + ⊂ H is a self-dual cone, i.e., its dual cone H + ∘ : = { ξ ∈ H ∣ ⟨ ξ , η ⟩ ≥ 0 , ∀ η ∈ H + } \mathcal{H}_{+}^{\circ}:=\{\xi\in\mathcal{H}\mid\langle\xi,\eta\rangle\ge0,\ \forall\eta\in\mathcal{H}_{+}\} H + ∘ := { ξ ∈ H ∣ ⟨ ξ , η ⟩ ≥ 0 , ∀ η ∈ H + } satisfies H + ∘ = H + \mathcal{H}_{+}^{\circ}=\mathcal{H}_{+} H + ∘ = H + ; and the following conditions hold:
J M J = M ′ ∩ B ( H ) JMJ=M'\cap B(\mathcal{H}) J M J = M ′ ∩ B ( H ) ;
J c J = c ∗ JcJ=c^{*} J c J = c ∗ for any c ∈ Z ( M ) c\in\mathcal{Z}(M) c ∈ Z ( M ) , the center of M M M ;
J ξ = ξ J\xi=\xi J ξ = ξ for any ξ ∈ H + \xi\in\mathcal{H}_{+} ξ ∈ H + ;
a ( J a J ) H + ⊂ H + a(JaJ)\mathcal{H}_{+}\subset\mathcal{H}_{+} a ( J a J ) H + ⊂ H + for any a ∈ M a\in M a ∈ M .
The standard form is unique up to unitary equivalence. Assume that ( M , H ~ , J ~ , H ~ + ) (M,\widetilde{\mathcal{H}},\widetilde{J},\widetilde{\mathcal{H}}_{+}) ( M , H , J , H + ) is another quadruple satisfying the above conditions. Then there exists a unique unitary u : H → H ~ u:\mathcal{H}\to\widetilde{\mathcal{H}} u : H → H such that
(a) π H ~ ( a ) = u π H ( a ) u ∗ \pi_{\widetilde{\mathcal{H}}}(a)=u\pi_{\mathcal{H}}(a)u^{*} π H ( a ) = u π H ( a ) u ∗ ( a ∈ M ) (a\in M) ( a ∈ M ) , where π H : M → B ( H ) \pi_{\mathcal{H}}:M\to B(\mathcal{H}) π H : M → B ( H ) and π H ~ : M → B ( H ~ ) \pi_{\widetilde{\mathcal{H}}}:M\to B(\widetilde{\mathcal{H}}) π H : M → B ( H ) are the representations of M M M on H \mathcal{H} H and H ~ \widetilde{\mathcal{H}} H , respectively; (b) J ~ = u J u ∗ \widetilde{J}=uJu^{*} J = u J u ∗ ; (c) H ~ + = u H + \widetilde{\mathcal{H}}_{+}=u\mathcal{H}_{+} H + = u H + .
Therefore, the standard form of M M M depends only on M M M itself, and we denote it by ( M , L 2 ( M ) , J M , L 2 ( M ) + ) (M,L^{2}(M),J_{M},L^{2}(M)_{+}) ( M , L 2 ( M ) , J M , L 2 ( M ) + ) . In particular, for any normal faithful state ϕ \phi ϕ on M M M , we can take
( M , L 2 ( M ) , J M , L 2 ( M ) + ) = ( M , L 2 ( M , ϕ ) , J ϕ , L 2 ( M , ϕ ) + ) , (M,L^{2}(M),J_{M},L^{2}(M)_{+})=(M,L^{2}(M,\phi),J_{\phi},L^{2}(M,\phi)_{+}), ( M , L 2 ( M ) , J M , L 2 ( M ) + ) = ( M , L 2 ( M , ϕ ) , J ϕ , L 2 ( M , ϕ ) + ) ,
where J ϕ J_{\phi} J ϕ is the modular conjugation operator determined by the polar decomposition S ϕ = J ϕ Δ ϕ 1 / 2 S_{\phi}=J_{\phi}\Delta_{\phi}^{1/2} S ϕ = J ϕ Δ ϕ 1/2 of the Tomita operator, and L 2 ( M , ϕ ) + L^{2}(M,\phi)_{+} L 2 ( M , ϕ ) + is the closure of { a ( J ϕ a J ϕ ) ξ ϕ ∣ a ∈ M } \{a(J_{\phi}aJ_{\phi})\xi_{\phi}\mid a\in M\} { a ( J ϕ a J ϕ ) ξ ϕ ∣ a ∈ M } [15 Lemma 2.9] .
By [15 Lemma 2.10] , any normal positive functional ω ∈ M ∗ + \omega\in M_{*}^{+} ω ∈ M ∗ + admits a unique representing vector ξ ω ∈ L 2 ( M ) + \xi_{\omega}\in L^{2}(M)_{+} ξ ω ∈ L 2 ( M ) + , i.e., ω = ⟨ ⋅ ξ ω , ξ ω ⟩ \omega=\langle\cdot\xi_{\omega},\xi_{\omega}\rangle ω = ⟨ ⋅ ξ ω , ξ ω ⟩ . Moreover, ξ ↦ ω ξ = ⟨ ⋅ ξ , ξ ⟩ \xi\mapsto\omega_{\xi}=\langle\cdot\xi,\xi\rangle ξ ↦ ω ξ = ⟨ ⋅ ξ , ξ ⟩ is a homeomorphism between L 2 ( M ) + L^{2}(M)_{+} L 2 ( M ) + and M ∗ + M_{*}^{+} M ∗ + .
Continuous cores and dual weights
For background on modular theory and continuous cores, we refer to [41 ] .
Let P P P be a countably decomposable von Neumann algebra, and let ρ \rho ρ be a normal faithful state on P P P . Let ξ ρ ∈ L 2 ( P ) + \xi_{\rho}\in L^{2}(P)_{+} ξ ρ ∈ L 2 ( P ) + be the vector representing ρ \rho ρ . The Tomita operator S ρ S_{\rho} S ρ is the closure on L 2 ( P ) L^{2}(P) L 2 ( P ) of the densely defined antilinear map x ξ ρ ↦ x ∗ ξ ρ x\xi_\rho\mapsto x^*\xi_\rho x ξ ρ ↦ x ∗ ξ ρ (x ∈ P x \in P x ∈ P ). Define the modular operator by Δ ρ = S ρ ∗ S ρ \Delta_\rho=S_\rho^*S_\rho Δ ρ = S ρ ∗ S ρ . The modular automorphism group σ ρ : R → Aut ( P ) \sigma^\rho:\mathbb{R}\to\operatorname{Aut}(P) σ ρ : R → Aut ( P ) is given by
σ t ρ ( x ) = Δ ρ i t x Δ ρ − i t ( x ∈ P , t ∈ R ) . \sigma_t^\rho(x)=\Delta_\rho^{it}x\Delta_\rho^{-it}\qquad(x\in P,\ t\in\mathbb{R}). σ t ρ ( x ) = Δ ρ i t x Δ ρ − i t ( x ∈ P , t ∈ R ) .
We say that x ∈ P x\in P x ∈ P is entire analytic for σ ρ \sigma^\rho σ ρ if t ↦ σ t ρ ( x ) t\mapsto\sigma_t^\rho(x) t ↦ σ t ρ ( x ) extends to a map z ↦ σ z ρ ( x ) z\mapsto\sigma_z^\rho(x) z ↦ σ z ρ ( x ) from C \mathbb{C} C to P P P such that z ↦ ω ( σ z ρ ( x ) ) z\mapsto\omega(\sigma_z^\rho(x)) z ↦ ω ( σ z ρ ( x )) is entire for every ω ∈ P ∗ \omega\in P_* ω ∈ P ∗ . Define the continuous core of P P P by
A = c ρ ( P ) = P ⋊ σ ρ R . \mathcal{A}=c_\rho(P)=P\rtimes_{\sigma^\rho}\mathbb{R}. A = c ρ ( P ) = P ⋊ σ ρ R .
We identify P P P with its canonical image in A \mathcal{A} A and denote by ( λ t ) t ∈ R (\lambda_t)_{t\in\mathbb{R}} ( λ t ) t ∈ R the canonical implementing unitaries, satisfying λ t x λ t ∗ = σ t ρ ( x ) \lambda_t x\lambda_t^*=\sigma_t^\rho(x) λ t x λ t ∗ = σ t ρ ( x ) (x ∈ P x\in P x ∈ P , t ∈ R t\in\mathbb{R} t ∈ R ). The dual action σ ^ : R ↷ A \widehat{\sigma}:\mathbb{R}\curvearrowright\mathcal{A} σ : R ↷ A is determined by
σ ^ r ( x ) = x , σ ^ r ( λ t ) = e − i r t λ t ( x ∈ P , r , t ∈ R ) . \widehat{\sigma}_r(x)=x,\qquad\widehat{\sigma}_r(\lambda_t)=e^{-irt}\lambda_t\qquad(x\in P,\ r,t\in\mathbb{R}). σ r ( x ) = x , σ r ( λ t ) = e − i r t λ t ( x ∈ P , r , t ∈ R ) .
The canonical operator-valued weight T : A + → P ^ + T:\mathcal{A}_+\to\widehat{P}_+ T : A + → P + is defined by
T ( x ) = ∫ R σ ^ r ( x ) d r ( x ∈ A + ) . T(x)=\int_{\mathbb{R}}\widehat{\sigma}_r(x)\,dr\qquad(x\in\mathcal{A}_+). T ( x ) = ∫ R σ r ( x ) d r ( x ∈ A + ) .
The dual weight of ρ \rho ρ is ψ A = ρ ∘ T \psi_\mathcal{A}=\rho\circ T ψ A = ρ ∘ T . Write τ A \tau_\mathcal{A} τ A for the canonical trace on A \mathcal{A} A , and let
h = d ψ A d τ A , H = log h , λ t = h i t = e i t H . h=\frac{d\psi_\mathcal{A}}{d\tau_\mathcal{A}},\qquad H=\log h,\qquad\lambda_t=h^{it}=e^{itH}. h = d τ A d ψ A , H = log h , λ t = h i t = e i t H .
Products of measurable operators are understood as closed products.
We next prove two lemmas that will be used in § 6 .
Lemma 2.1. For every positive integrable function g g g on R \mathbb{R} R ,
ψ A ( g ( H ) ) = ρ ( T ( g ( H ) ) ) = ∫ R g ( s ) d s . (5) \psi_\mathcal{A}(g(H))=\rho(T(g(H)))=\int_{\mathbb{R}}g(s)\,ds. \tag*{(5)} ψ A ( g ( H )) = ρ ( T ( g ( H ))) = ∫ R g ( s ) d s . (5) In particular, we have
τ A ( 1 ( s , ∞ ) ( h ) ) = 1 s ( s > 0 ) . (6) \tau_\mathcal{A}(\mathbf{1}_{(s,\infty)}(h))=\frac{1}{s}\qquad(s>0). \tag*{(6)} τ A ( 1 ( s , ∞ ) ( h )) = s 1 ( s > 0 ) . (6) Proof. For r , t ∈ R r,t\in\mathbb{R} r , t ∈ R , by functional calculus we have
e i t σ ^ r ( H ) = σ ^ r ( e i t H ) = σ ^ r ( λ t ) = e − i r t e i t H = e i t ( H − r ) . e^{it\widehat{\sigma}_r(H)}=\widehat{\sigma}_r(e^{itH})=\widehat{\sigma}_r(\lambda_t)=e^{-irt}e^{itH}=e^{it(H-r)}. e i t σ r ( H ) = σ r ( e i t H ) = σ r ( λ t ) = e − i r t e i t H = e i t ( H − r ) . By uniqueness of the self-adjoint generator of a strongly continuous unitary group,
σ ^ r ( H ) = H − r . (7) \widehat{\sigma}_r(H)=H-r. \tag*{(7)} σ r ( H ) = H − r . (7) Let E H E_H E H be the spectral measure of H H H . For a positive integrable function g g g , by (7) we have
T ( g ( H ) ) = ∫ R g ( H − r ) d r = ∫ R ( ∫ R g ( s − r ) d r ) d E H ( s ) = ( ∫ R g ( s ) d s ) 1. (8) T(g(H))=\int_{\mathbb{R}}g(H-r)\,dr=\int_{\mathbb{R}}\left(\int_{\mathbb{R}}g(s-r)\,dr\right)dE_H(s)=\left(\int_{\mathbb{R}}g(s)\,ds\right)1. \tag*{(8)} T ( g ( H )) = ∫ R g ( H − r ) d r = ∫ R ( ∫ R g ( s − r ) d r ) d E H ( s ) = ( ∫ R g ( s ) d s ) 1. (8) Since ψ A = ρ ∘ T \psi_\mathcal{A}=\rho\circ T ψ A = ρ ∘ T and ρ ( 1 ) = 1 \rho(1)=1 ρ ( 1 ) = 1 , by (8) we have (5) .
For s > 0 s>0 s > 0 , by the definition of h h h and (5) , we have
τ A ( 1 ( s , ∞ ) ( h ) ) = ψ A ( h − 1 1 ( s , ∞ ) ( h ) ) = ψ A ( e − H 1 ( log s , ∞ ) ( H ) ) = ∫ log s ∞ e − t d t = 1 s . \tau_\mathcal{A}(\mathbf{1}_{(s,\infty)}(h))=\psi_\mathcal{A}(h^{-1}\mathbf{1}_{(s,\infty)}(h))=\psi_\mathcal{A}(e^{-H}\mathbf{1}_{(\log s,\infty)}(H))=\int_{\log s}^{\infty}e^{-t}\,dt=\frac{1}{s}. τ A ( 1 ( s , ∞ ) ( h )) = ψ A ( h − 1 1 ( s , ∞ ) ( h )) = ψ A ( e − H 1 ( l o g s , ∞ ) ( H )) = ∫ l o g s ∞ e − t d t = s 1 . See also [17 Theorem 1.2(2)] . In particular, h h h is τ A \tau_\mathcal{A} τ A -measurable. □ \square □
Lemma 2.2. For every 0 ≠ d ∈ A + 0\ne d\in\mathcal{A}_+ 0 = d ∈ A + , there exists x ∈ A + x\in\mathcal{A}_+ x ∈ A + such that
0 ≠ x ≤ d , ψ A ( x ) < ∞ . 0\ne x\le d,\qquad\psi_\mathcal{A}(x)<\infty. 0 = x ≤ d , ψ A ( x ) < ∞.
Proof. Take δ > 0 \delta> 0 δ > 0 such that 1 [ δ , ∞ ) ( d ) ≠ 0 \mathbf{1}_{[\delta,\infty)}(d) \ne0 1 [ δ , ∞ ) ( d ) = 0 . By semifiniteness of τ A \tau_A τ A , there is a nonzero projection
p ≤ 1 [ δ , ∞ ) ( d ) , 0 < τ A ( p ) < ∞ . p \le\mathbf{1}_{[\delta,\infty)}(d), \qquad0 < \tau_A(p) < \infty. p ≤ 1 [ δ , ∞ ) ( d ) , 0 < τ A ( p ) < ∞.
By (6) , take s > 0 s > 0 s > 0 such that τ A ( 1 ( s , ∞ ) ( h ) ) < τ A ( p ) \tau_A(\mathbf{1}_{(s,\infty)}(h)) < \tau_A(p) τ A ( 1 ( s , ∞ ) ( h )) < τ A ( p ) , and let e = 1 [ 0 , s ] ( h ) e = \mathbf{1}_{[0,s]}(h) e = 1 [ 0 , s ] ( h ) and q = p ∧ e q = p \wedge e q = p ∧ e . Let ( 1 − e ) p = v ∣ ( 1 − e ) p ∣ (1-e)p = v\lvert(1-e)p\rvert ( 1 − e ) p = v ∣( 1 − e ) p ∣ be the polar decomposition. The kernel of ( 1 − e ) p (1-e)p ( 1 − e ) p inside the range of p p p is the range of q = p ∧ e q = p \wedge e q = p ∧ e , and the range of ( 1 − e ) p (1-e)p ( 1 − e ) p is contained in the range of 1 − e 1-e 1 − e . Hence
v ∗ v = p − q , v v ∗ ≤ 1 − e . v^*v = p-q, \qquad vv^* \le1-e. v ∗ v = p − q , v v ∗ ≤ 1 − e .
Hence p − q ∼ v v ∗ ≤ 1 − e p-q \sim vv^* \le1-e p − q ∼ v v ∗ ≤ 1 − e , so p − q ≾ 1 − e p-q \precsim1-e p − q ≾ 1 − e and
τ A ( q ) ≥ τ A ( p ) − τ A ( 1 − e ) > 0. \tau_A(q) \ge\tau_A(p) - \tau_A(1-e) > 0. τ A ( q ) ≥ τ A ( p ) − τ A ( 1 − e ) > 0.
Since q ≤ e q \le e q ≤ e , we have
ψ A ( q ) = τ A ( q h q ) = τ A ( q ( e h e ) q ) ≤ s τ A ( q ) < ∞ . \psi_A(q) = \tau_A(qhq) = \tau_A(q(ehe)q) \le s\tau_A(q) < \infty. ψ A ( q ) = τ A ( q h q ) = τ A ( q ( e h e ) q ) ≤ s τ A ( q ) < ∞.
Hence x = δ q x = \delta q x = δ q satisfies
0 ≠ x = δ q ≤ δ p ≤ d , ψ A ( x ) = δ ψ A ( q ) < ∞ . 0 \ne x = \delta q \le\delta p \le d, \qquad\psi_A(x) = \delta\psi_A(q) < \infty. 0 = x = δ q ≤ δ p ≤ d , ψ A ( x ) = δ ψ A ( q ) < ∞.
Centralizers and bicentralizers
Let M M M be a countably decomposable von Neumann algebra, and let ϕ \phi ϕ be a normal faithful state on M M M . The centralizer of ϕ \phi ϕ is
M ϕ = { a ∈ M ∣ ϕ ( a x ) = ϕ ( x a ) , ∀ x ∈ M } = { a ∈ M ∣ σ t ϕ ( a ) = a , ∀ t ∈ R } . M_\phi= \{a \in M \mid\phi(ax) = \phi(xa),\ \forall x \in M\} = \{a \in M \mid\sigma_t^\phi(a) = a,\ \forall t \in\mathbb{R}\}. M ϕ = { a ∈ M ∣ ϕ ( a x ) = ϕ ( x a ) , ∀ x ∈ M } = { a ∈ M ∣ σ t ϕ ( a ) = a , ∀ t ∈ R } .
The asymptotic centralizer of ϕ \phi ϕ is
AC ( M , ϕ ) = { ( a n ) ∈ ℓ ∞ ( N , M ) ∣ ∥ a n ϕ − ϕ a n ∥ M ∗ → 0 } , \operatorname{AC}(M,\phi) = \{(a_n) \in\ell^\infty(\mathbb{N},M) \mid\|a_n\phi-\phi a_n\|_{M_*} \to0\}, AC ( M , ϕ ) = {( a n ) ∈ ℓ ∞ ( N , M ) ∣ ∥ a n ϕ − ϕ a n ∥ M ∗ → 0 } ,
where ( a ϕ ) ( x ) = ϕ ( x a ) (a\phi)(x) = \phi(xa) ( a ϕ ) ( x ) = ϕ ( x a ) and ( ϕ a ) ( x ) = ϕ ( a x ) (\phi a)(x) = \phi(ax) ( ϕ a ) ( x ) = ϕ ( a x ) for a , x ∈ M a,x \in M a , x ∈ M . For a W ∗ W^* W ∗ -inclusion M ⊂ P M \subset P M ⊂ P , the relative bicentralizer is defined by
B ( M ⊂ P , ϕ ) = { x ∈ P ∣ [ x , a n ] ⟶ s 0 , ∀ ( a n ) ∈ AC ( M , ϕ ) } , \operatorname{B}(M \subset P,\phi) = \{x \in P \mid[x,a_n] \mathop{\longrightarrow}^{s} 0,\ \forall(a_n) \in\operatorname{AC}(M,\phi)\}, B ( M ⊂ P , ϕ ) = { x ∈ P ∣ [ x , a n ] ⟶ s 0 , ∀ ( a n ) ∈ AC ( M , ϕ )} ,
where s s s denotes strong operator convergence, see [AHHM20, Definition 3.1]. For P = M P=M P = M , we have the bicentralizer of ϕ \phi ϕ
B ( M , ϕ ) = B ( M ⊂ M , ϕ ) . \operatorname{B}(M,\phi) = \operatorname{B}(M \subset M,\phi). B ( M , ϕ ) = B ( M ⊂ M , ϕ ) .
Marrakchi’s relative bicentralizer conjecture [32 Conjecture C] asserts that, for every inclusion M ⊂ P M \subset P M ⊂ P of countably decomposable von Neumann algebras with a normal faithful conditional expectation E : P → M E:P \to M E : P → M and every normal faithful state ϕ \phi ϕ on M M M ,
B ( M ⊂ c ( P ) , ϕ ) = { ϕ i t ∣ t ∈ R } ′ ′ ∨ ( M ′ ∩ c ( P ) ) . (9) \operatorname{B}(M \subset c(P),\phi) = \{\phi^{it} \mid t \in\mathbb{R}\}^{\prime\prime} \vee(M' \cap c(P)). \tag*{(9)} B ( M ⊂ c ( P ) , ϕ ) = { ϕ i t ∣ t ∈ R } ′′ ∨ ( M ′ ∩ c ( P )) . (9)
Here we use the compatible realizations c ϕ ( M ) ⊂ c ϕ ∘ E ( P ) = c ( P ) c_\phi(M) \subset c_{\phi\circ E}(P) = c(P) c ϕ ( M ) ⊂ c ϕ ∘ E ( P ) = c ( P ) , and ϕ i t \phi^{it} ϕ i t denotes the canonical implementing unitary in c ϕ ( M ) c_\phi(M) c ϕ ( M ) .
Poisson boundaries of general von Neumann algebras
In this section, we develop the Poisson boundary theory for general von Neumann algebras, following the framework of Das–Peterson [9 ] for tracial von Neumann algebras. The proofs of ergodicity, double ergodicity, and rigidity are in the same spirit as those of [9 ] . However, the coefficient families ( z k ) (z_k) ( z k ) and ( y k ) (y_k) ( y k ) for P ϕ \mathcal{P}_\phi P ϕ and Q ϕ \mathcal{Q}_\phi Q ϕ need not coincide in the nontracial case, whereas we can take y k = z k y_k = z_k y k = z k if ϕ ∣ M \phi|_M ϕ ∣ M is tracial (Remark 3.7). Hence, in the general setting, we cannot recover all the results proved in [9 ] for finite von Neumann algebras.
Throughout this section, we fix a von Neumann algebra M M M with separable predual.
Definition 3.1. Let M ⊂ A M \subset A M ⊂ A be a W ∗ W^* W ∗ -inclusion. A normal state ϕ \phi ϕ on A A A is called an M M M -hyperstate if its restriction ϕ ∣ M \phi|_M ϕ ∣ M is faithful. Given a fixed normal faithful state φ \varphi φ on M M M , we say that ϕ \phi ϕ is an ( M , φ ) (M,\varphi) ( M , φ ) -hyperstate if ϕ ∣ M = φ \phi|_M = \varphi ϕ ∣ M = φ .
The framework of Das–Peterson [9 ] concerns ( M , τ ) (M,\tau) ( M , τ ) -hyperstates for tracial von Neumann algebras ( M , τ ) (M,\tau) ( M , τ ) , whereas we work with general M M M -hyperstates without fixing a normal faithful state ϕ \phi ϕ on M M M in advance.
The following proposition associates with each normal M M M -hyperstate on B ( L 2 ( M ) ) B(L^2(M)) B ( L 2 ( M )) an M M M -bimodular ucp map, which we use to define the Poisson boundary.
Proposition 3.2. Let φ ∈ B ( L 2 ( M ) ) ∗ \varphi\in B(L^2(M))_\ast φ ∈ B ( L 2 ( M ) ) ∗ be a normal M M M -hyperstate. Set ϕ = φ ∣ M \phi=\varphi|_M ϕ = φ ∣ M , and let ξ ϕ ∈ L 2 ( M ) + \xi_\phi\in L^2(M)_+ ξ ϕ ∈ L 2 ( M ) + be the canonical cyclic and separating vector implementing ϕ \phi ϕ . Then there exists a unique normal M M M -bimodular ucp map P φ : B ( L 2 ( M ) ) → B ( L 2 ( M ) ) P_\varphi:B(L^2(M))\to B(L^2(M)) P φ : B ( L 2 ( M )) → B ( L 2 ( M )) such that for all T ∈ B ( L 2 ( M ) ) T\in B(L^2(M)) T ∈ B ( L 2 ( M )) ,
φ ( T ) = ⟨ P φ ( T ) ξ ϕ , ξ ϕ ⟩ . \varphi(T)=\langle P_\varphi(T)\xi_\phi,\xi_\phi\rangle. φ ( T ) = ⟨ P φ ( T ) ξ ϕ , ξ ϕ ⟩ . Moreover, there exists a (not necessarily unique) sequence ( z k ) k = 1 ∞ ⊂ M (z_k)_{k=1}^{\infty}\subset M ( z k ) k = 1 ∞ ⊂ M such that
P φ ( T ) = ∑ k = 1 ∞ ( J z k ∗ J ) T ( J z k J ) ( T ∈ B ( L 2 ( M ) ) ) , P_\varphi(T)=\sum_{k=1}^{\infty}(Jz_k^*J)T(Jz_kJ)\qquad(T\in B(L^2(M))), P φ ( T ) = k = 1 ∑ ∞ ( J z k ∗ J ) T ( J z k J ) ( T ∈ B ( L 2 ( M ))) , where J J J is the modular conjugation operator on L 2 ( M ) L^2(M) L 2 ( M ) .
Proof. We first prove the existence and construct the explicit form of P φ P_\varphi P φ .
Since φ \varphi φ is normal, there exists a sequence ( ξ k ) ⊂ L 2 ( M ) (\xi_k)\subset L^2(M) ( ξ k ) ⊂ L 2 ( M ) such that
φ ( T ) = ∑ k ⟨ T ξ k , ξ k ⟩ ( T ∈ B ( L 2 ( M ) ) ) . \varphi(T)=\sum_k\langle T\xi_k,\xi_k\rangle\qquad(T\in B(L^2(M))). φ ( T ) = k ∑ ⟨ T ξ k , ξ k ⟩ ( T ∈ B ( L 2 ( M ))) .
For any x ∈ M x\in M x ∈ M , since φ ∣ M = ϕ \varphi|_M=\phi φ ∣ M = ϕ , we have
∥ x ξ ϕ ∥ 2 = ϕ ( x ∗ x ) = φ ( x ∗ x ) = ∑ k ∥ x ξ k ∥ 2 . \|x\xi_\phi\|^2=\phi(x^*x)=\varphi(x^*x)=\sum_k\|x\xi_k\|^2. ∥ x ξ ϕ ∥ 2 = ϕ ( x ∗ x ) = φ ( x ∗ x ) = k ∑ ∥ x ξ k ∥ 2 .
Hence
∥ x ξ n ∥ 2 ≤ ∥ x ξ ϕ ∥ 2 ( x ∈ M , n ≥ 1 ) . (10) \|x\xi_n\|^2\leq\|x\xi_\phi\|^2\qquad(x\in M,\ n\geq1). \tag*{(10)} ∥ x ξ n ∥ 2 ≤ ∥ x ξ ϕ ∥ 2 ( x ∈ M , n ≥ 1 ) . (10)
Define the operator R n : M ξ ϕ → L 2 ( M ) R_n:M\xi_\phi\to L^2(M) R n : M ξ ϕ → L 2 ( M ) by R n ( x ξ ϕ ) = x ξ n R_n(x\xi_\phi)=x\xi_n R n ( x ξ ϕ ) = x ξ n . By (10) , ∥ R n ( x ξ ϕ ) ∥ ≤ ∥ x ξ ϕ ∥ \|R_n(x\xi_\phi)\|\leq\|x\xi_\phi\| ∥ R n ( x ξ ϕ ) ∥ ≤ ∥ x ξ ϕ ∥ , so R n R_n R n extends to a contraction on L 2 ( M ) L^2(M) L 2 ( M ) . Moreover, for any x , y ∈ M x,y\in M x , y ∈ M , we have R n ( y x ξ ϕ ) = y x ξ n = y R n ( x ξ ϕ ) R_n(yx\xi_\phi)=yx\xi_n=yR_n(x\xi_\phi) R n ( y x ξ ϕ ) = y x ξ n = y R n ( x ξ ϕ ) . Hence R n ∈ M ′ R_n\in M' R n ∈ M ′ . By standard Tomita–Takesaki theory, M ′ = J M J M'=JMJ M ′ = J M J , so there exists z n ∈ M z_n\in M z n ∈ M such that R n = J z n J R_n=Jz_nJ R n = J z n J .
Now, for any x , y ∈ M x,y\in M x , y ∈ M and T ∈ B ( L 2 ( M ) ) T\in B(L^2(M)) T ∈ B ( L 2 ( M )) , we calculate
φ ( y ∗ T x ) = ∑ k ⟨ T x ξ k , y ξ k ⟩ = ∑ k ⟨ T R k ( x ξ ϕ ) , R k ( y ξ ϕ ) ⟩ = ∑ k ⟨ T J z k J x ξ ϕ , J z k J y ξ ϕ ⟩ = ⟨ ( ∑ k J z k ∗ J T J z k J ) x ξ ϕ , y ξ ϕ ⟩ . (11) \begin{aligned}
\varphi(y^*Tx)
&=\sum_k\langle Tx\xi_k,y\xi_k\rangle
=\sum_k\langle TR_k(x\xi_\phi),R_k(y\xi_\phi)\rangle\\
&=\sum_k\langle TJz_kJx\xi_\phi,Jz_kJy\xi_\phi\rangle\\
&=\left\langle\left(\sum_kJz_k^*JTJz_kJ\right)x\xi_\phi,y\xi_\phi\right\rangle. \tag*{(11)}
\end{aligned} φ ( y ∗ T x ) = k ∑ ⟨ T x ξ k , y ξ k ⟩ = k ∑ ⟨ T R k ( x ξ ϕ ) , R k ( y ξ ϕ )⟩ = k ∑ ⟨ T J z k J x ξ ϕ , J z k J y ξ ϕ ⟩ = ⟨ ( k ∑ J z k ∗ J T J z k J ) x ξ ϕ , y ξ ϕ ⟩ . (11)
Let P φ ( T ) = ∑ k ( J z k ∗ J ) T ( J z k J ) P_\varphi(T)=\sum_k(Jz_k^*J)T(Jz_kJ) P φ ( T ) = ∑ k ( J z k ∗ J ) T ( J z k J ) . By (11) , φ ( y ∗ T x ) = ⟨ P φ ( T ) x ξ ϕ , y ξ ϕ ⟩ \varphi(y^*Tx)=\langle P_\varphi(T)x\xi_\phi,y\xi_\phi\rangle φ ( y ∗ T x ) = ⟨ P φ ( T ) x ξ ϕ , y ξ ϕ ⟩ . Taking x = y = 1 x=y=1 x = y = 1 , we immediately obtain φ ( T ) = ⟨ P φ ( T ) ξ ϕ , ξ ϕ ⟩ \varphi(T)=\langle P_\varphi(T)\xi_\phi,\xi_\phi\rangle φ ( T ) = ⟨ P φ ( T ) ξ ϕ , ξ ϕ ⟩ .
Since M ξ ϕ M\xi_\phi M ξ ϕ is dense in L 2 ( M ) L^2(M) L 2 ( M ) , the identity
⟨ P φ ( T ) x ξ ϕ , y ξ ϕ ⟩ = φ ( y ∗ T x ) \langle P_\varphi(T)x\xi_\phi,y\xi_\phi\rangle=\varphi(y^*Tx) ⟨ P φ ( T ) x ξ ϕ , y ξ ϕ ⟩ = φ ( y ∗ T x )
uniquely determines P φ ( T ) P_\varphi(T) P φ ( T ) for every T T T . Hence P φ P_\varphi P φ is unique.
Taking T = 1 T=1 T = 1 in (11) and using φ ∣ M = ϕ \varphi|_M=\phi φ ∣ M = ϕ , we have ∑ k z k ∗ z k = 1 \sum_k z_k^*z_k=1 ∑ k z k ∗ z k = 1 . Since J z k J ∈ M ′ Jz_kJ\in M' J z k J ∈ M ′ , P φ P_\varphi P φ is a normal M M M -bimodular ucp map. □
Definition 3.3. Let φ \varphi φ be a normal M M M -hyperstate on B ( L 2 ( M ) ) B(L^2(M)) B ( L 2 ( M )) . Let Har ( P φ ) = { T ∈ B ( L 2 ( M ) ) ∣ P φ ( T ) = T } \operatorname{Har}(\mathcal{P}_{\varphi}) = \{T \in B(L^2(M)) \mid\mathcal{P}_{\varphi}(T) = T\} Har ( P φ ) = { T ∈ B ( L 2 ( M )) ∣ P φ ( T ) = T } denote the space of φ \varphi φ -harmonic operators. The Poisson boundary ( B φ , ζ ) (\mathcal{B}_{\varphi},\zeta) ( B φ , ζ ) is defined as a von Neumann algebra that is isomorphic to Har ( P φ ) \operatorname{Har}(\mathcal{P}_{\varphi}) Har ( P φ ) as an operator system, together with the canonical hyperstate ζ : = φ ∣ Har ( P φ ) \zeta:= \left.\varphi\right|_{\operatorname{Har}(\mathcal{P}_{\varphi})} ζ := φ ∣ Har ( P φ ) .
The next proposition shows that the Poisson boundary is well defined and amenable.
Proposition 3.4. Let φ \varphi φ be a normal M M M -hyperstate on B ( L 2 ( M ) ) B(L^2(M)) B ( L 2 ( M )) . The Poisson boundary B φ \mathcal{B}_{\varphi} B φ always exists, is unique, and is amenable.
Proof. Let ω ∈ β N ∖ N \omega\in\beta\mathbb{N}\setminus\mathbb{N} ω ∈ β N ∖ N be a free ultrafilter, and let C n = ( 1 / n ) ∑ k = 0 n − 1 P φ k C_n = (1/n)\sum_{k=0}^{n-1}\mathcal{P}_{\varphi}^{k} C n = ( 1/ n ) ∑ k = 0 n − 1 P φ k . We define a conditional expectation E : B ( L 2 ( M ) ) → Har ( P φ ) E:B(L^2(M))\to\operatorname{Har}(\mathcal{P}_{\varphi}) E : B ( L 2 ( M )) → Har ( P φ ) by
E ( T ) = w o - l i m n → ω C n ( T ) ( T ∈ B ( L 2 ( M ) ) ) . E(T)=\mathop{\mathrm{wo\text{-}lim}}_{n\to\omega} C_n(T) \quad(T\in B(L^2(M))). E ( T ) = wo - lim n → ω C n ( T ) ( T ∈ B ( L 2 ( M ))) .
By the Choi–Effros theorem [5 ] , equipping the operator system Har ( P φ ) \operatorname{Har}(\mathcal{P}_{\varphi}) Har ( P φ ) with the Choi–Effros product defined by
x ∘ y = E ( x y ) ( x , y ∈ Har ( P φ ) ) x\circ y=E(xy) \quad(x,y\in\operatorname{Har}(\mathcal{P}_{\varphi})) x ∘ y = E ( x y ) ( x , y ∈ Har ( P φ ))
makes ( Har ( P φ ) , ∘ ) (\operatorname{Har}(\mathcal{P}_{\varphi}),\circ) ( Har ( P φ ) , ∘ ) a C ∗ C^* C ∗ -algebra with its original norm. The harmonic operator system is ultraweakly closed, hence this C ∗ C^* C ∗ -algebra is a von Neumann algebra, with its canonical ultraweak topology agreeing with the inherited one, see also [24 ] . The inclusion M ⊂ Har ( P φ ) M\subset\operatorname{Har}(\mathcal{P}_{\varphi}) M ⊂ Har ( P φ ) is multiplicative for this product, since E E E is M M M -bimodular.
The uniqueness of B φ \mathcal{B}_{\varphi} B φ follows from the fact that any two von Neumann algebras that are completely isometrically isomorphic as operator systems are necessarily ∗ * ∗ -isomorphic.
Finally, since B φ ≅ Har ( P φ ) \mathcal{B}_{\varphi}\cong\operatorname{Har}(\mathcal{P}_{\varphi}) B φ ≅ Har ( P φ ) as injective operator systems, B φ \mathcal{B}_{\varphi} B φ is amenable as a von Neumann algebra. □
Under the following symmetry condition, we can also represent φ \varphi φ by an M M M -inner ucp map.
Definition 3.5. We say that a normal M M M -hyperstate φ ∈ B ( L 2 ( M ) ) ∗ \varphi\in B(L^2(M))_* φ ∈ B ( L 2 ( M ) ) ∗ is regular if φ ( J z ∗ J ) = φ ( z ) \varphi(Jz^*J)=\varphi(z) φ ( J z ∗ J ) = φ ( z ) for all z ∈ M z\in M z ∈ M .
For regular hyperstates, Proposition 3.2 has the following counterpart, with M M M and J M J JMJ J M J interchanged.
Proposition 3.6. Let φ ∈ B ( L 2 ( M ) ) ∗ \varphi\in B(L^2(M))_* φ ∈ B ( L 2 ( M ) ) ∗ be a normal regular M M M -hyperstate, and let ϕ = φ ∣ M \phi=\left.\varphi\right|_M ϕ = φ ∣ M . Then there exists a sequence ( y k ) k = 1 ∞ ⊂ M (y_k)_{k=1}^{\infty}\subset M ( y k ) k = 1 ∞ ⊂ M with ∑ k y k y k ∗ = 1 \sum_k y_k y_k^*=1 ∑ k y k y k ∗ = 1 such that
Q φ ( T ) = ∑ k y k T y k ∗ ( T ∈ B ( L 2 ( M ) ) ) Q_{\varphi}(T)=\sum_k y_k T y_k^* \quad(T\in B(L^2(M))) Q φ ( T ) = k ∑ y k T y k ∗ ( T ∈ B ( L 2 ( M ))) defines a normal J M J JMJ J M J -bimodular ucp map satisfying
φ ( T ) = ⟨ Q φ ( T ) ξ ϕ , ξ ϕ ⟩ ( T ∈ B ( L 2 ( M ) ) ) . \varphi(T)=\langle Q_{\varphi}(T)\xi_{\phi},\xi_{\phi}\rangle\quad(T\in B(L^2(M))). φ ( T ) = ⟨ Q φ ( T ) ξ ϕ , ξ ϕ ⟩ ( T ∈ B ( L 2 ( M ))) .
Proof. Let φ ∘ = φ ∣ J M J \varphi^{\circ}=\left.\varphi\right|_{JMJ} φ ∘ = φ ∣ J M J . Since φ \varphi φ is regular and J ξ ϕ = ξ ϕ J\xi_{\phi}=\xi_{\phi} J ξ ϕ = ξ ϕ , we have
⟨ J z ∗ J ξ ϕ , ξ ϕ ⟩ = ϕ ( z ) = φ ∘ ( J z ∗ J ) ( z ∈ M ) . \langle Jz^*J\xi_{\phi},\xi_{\phi}\rangle=\phi(z)=\varphi^{\circ}(Jz^*J) \quad(z\in M). ⟨ J z ∗ J ξ ϕ , ξ ϕ ⟩ = ϕ ( z ) = φ ∘ ( J z ∗ J ) ( z ∈ M ) .
Hence ξ ϕ \xi_{\phi} ξ ϕ is the canonical cyclic and separating vector implementing φ ∘ \varphi^{\circ} φ ∘ . By applying Proposition 3.2 to J M J JMJ J M J , we have the required map and coefficients. □
Definition 3.8. Let φ \varphi φ be a normal regular M M M -hyperstate. We say that φ \varphi φ is strongly generating if there exist sequences ( z k ) , ( y k ) ⊂ M (z_k),(y_k)\subset M ( z k ) , ( y k ) ⊂ M with
C [ ( z k ) ] ‾ w = C [ ( y k ) ] ‾ w = M \overline{\mathbb{C}[(z_k)]}^{\,w}=\overline{\mathbb{C}[(y_k)]}^{\,w}=M C [( z k )] w = C [( y k )] w = M
such that P φ P_\varphi P φ and Q φ Q_\varphi Q φ have the standard forms
P φ ( T ) = ∑ k ( J z k ∗ J ) T ( J z k J ) , Q φ ( T ) = ∑ k y k T y k ∗ ( T ∈ B ( L 2 ( M ) ) ) . P_\varphi(T)=\sum_k (Jz_k^*J)T(Jz_kJ),\qquad Q_\varphi(T)=\sum_k y_kTy_k^* \qquad(T\in B(L^2(M))). P φ ( T ) = k ∑ ( J z k ∗ J ) T ( J z k J ) , Q φ ( T ) = k ∑ y k T y k ∗ ( T ∈ B ( L 2 ( M ))) .
Here C [ ( z k ) ] \mathbb{C}[(z_k)] C [( z k )] and C [ ( y k ) ] \mathbb{C}[(y_k)] C [( y k )] are the unital subalgebras (rather than ∗ \ast ∗ -algebras) generated by ( z k ) (z_k) ( z k ) and ( y k ) (y_k) ( y k ) , respectively. We abbreviate normal regular strongly generating hyperstate as nrsg hyperstate. In the nontracial setting we assume both density conditions. We do not assert that one follows from the other.
We now prove the existence of nrsg hyperstates. Although regularity requires φ ( J T ∗ J ) = φ ( T ) \varphi(JT^*J)=\varphi(T) φ ( J T ∗ J ) = φ ( T ) only for T ∈ M T\in M T ∈ M , it is easier to impose this symmetry for all T ∈ B ( L 2 ( M ) ) T\in B(L^2(M)) T ∈ B ( L 2 ( M )) . Indeed, for any normal faithful state φ 0 \varphi_0 φ 0 on B ( L 2 ( M ) ) B(L^2(M)) B ( L 2 ( M )) , the state φ ( T ) = 1 2 ( φ 0 ( T ) + φ 0 ( J T ∗ J ) ) \varphi(T)=\frac{1}{2}(\varphi_0(T)+\varphi_0(JT^*J)) φ ( T ) = 2 1 ( φ 0 ( T ) + φ 0 ( J T ∗ J )) ( T ∈ B ( L 2 ( M ) ) ) (T\in B(L^2(M))) ( T ∈ B ( L 2 ( M ))) is a normal faithful symmetric M M M -hyperstate, and hence regular.
Fix a normal faithful state φ \varphi φ on M M M , and let ξ = ξ φ ∈ L 2 ( M ) + \xi=\xi_\varphi\in L^2(M)_+ ξ = ξ φ ∈ L 2 ( M ) + . To construct a strongly generating symmetric ( M , φ ) (M,\varphi) ( M , φ ) -hyperstate, we want the coefficient families ( z k ) (z_k) ( z k ) and ( y k ) (y_k) ( y k ) in Definition 3.8 to generate M M M and satisfy
∑ k ⟨ T y k ∗ ξ , y k ∗ ξ ⟩ = ∑ k ⟨ T J y k ∗ ξ , J y k ∗ ξ ⟩ ( T ∈ B ( L 2 ( M ) ) ) . \sum_k\langle Ty_k^*\xi,y_k^*\xi\rangle=\sum_k\langle TJy_k^*\xi,Jy_k^*\xi\rangle\qquad(T\in B(L^2(M))). k ∑ ⟨ T y k ∗ ξ , y k ∗ ξ ⟩ = k ∑ ⟨ T J y k ∗ ξ , J y k ∗ ξ ⟩ ( T ∈ B ( L 2 ( M ))) .
This equality holds if ( J y k ∗ ξ ) (Jy_k^*\xi) ( J y k ∗ ξ ) is a permutation of ( y k ∗ ξ ) (y_k^*\xi) ( y k ∗ ξ ) . For analytic coefficients, we have
J y k ∗ ξ = σ − i / 2 φ ( y k ) ξ = ( σ i / 2 φ ( y k ∗ ) ) ∗ ξ . Jy_k^*\xi=\sigma^\varphi_{-i/2}(y_k)\xi=(\sigma^\varphi_{i/2}(y_k^*))^*\xi. J y k ∗ ξ = σ − i /2 φ ( y k ) ξ = ( σ i /2 φ ( y k ∗ ) ) ∗ ξ .
Hence we naturally ask that y k ↦ σ i / 2 φ ( y k ∗ ) y_k\mapsto\sigma^\varphi_{i/2}(y_k^*) y k ↦ σ i /2 φ ( y k ∗ ) permute the coefficient family. We therefore take an analytic family ( a n ) (a_n) ( a n ) with weakly dense linear span and let b n = σ − i / 2 φ ( a n ∗ ) b_n=\sigma^\varphi_{-i/2}(a_n^*) b n = σ − i /2 φ ( a n ∗ ) . We pair the coefficients c n a n ∗ \sqrt{c_n}a_n^* c n a n ∗ and c n b n ∗ \sqrt{c_n}b_n^* c n b n ∗ , using the same weight c n > 0 c_n>0 c n > 0 .
It is not easy to arrange φ ∘ Q 0 = φ \varphi\circ Q_0=\varphi φ ∘ Q 0 = φ directly for Q 0 ( z ) = ∑ n c n ( a n ∗ z a n + b n ∗ z b n ) Q_0(z)=\sum_n c_n(a_n^*za_n+b_n^*zb_n) Q 0 ( z ) = ∑ n c n ( a n ∗ z a n + b n ∗ z b n ) ( z ∈ M ) (z\in M) ( z ∈ M ) with these paired coefficients. However, we can take the weights c n c_n c n small enough that ψ : = φ ∘ Q 0 ≤ φ \psi:=\varphi\circ Q_0\leq\varphi ψ := φ ∘ Q 0 ≤ φ and ∑ n c n ( ∥ a n ∥ ∞ 2 + ∥ b n ∥ ∞ 2 ) < ∞ \sum_n c_n(\|a_n\|_\infty^2+\|b_n\|_\infty^2)<\infty ∑ n c n ( ∥ a n ∥ ∞ 2 + ∥ b n ∥ ∞ 2 ) < ∞ . By [15 Lemma 2.10(1)] , let η ∈ L 2 ( M ) + \eta\in L^2(M)_+ η ∈ L 2 ( M ) + represent φ − ψ \varphi-\psi φ − ψ . Since J η = η J\eta=\eta J η = η , adding the positive functional ⟨ ⋅ η , η ⟩ \langle\mathord\cdot\,\eta,\eta\rangle ⟨ ⋅ η , η ⟩ to the paired sum preserves this symmetry and makes its restriction to M M M equal to φ \varphi φ .
The following example gives this construction. The properties in (ii) follow naturally from it and will be used in § 6 .
Example 3.9. Let φ \varphi φ be a normal faithful state on M M M , and write ξ = ξ φ ∈ L 2 ( M ) + \xi=\xi_\varphi\in L^2(M)_+ ξ = ξ φ ∈ L 2 ( M ) + . There exists a faithful nrsg ( M , φ ) (M,\varphi) ( M , φ ) -hyperstate φ \varphi φ on B ( L 2 ( M ) ) B(L^2(M)) B ( L 2 ( M )) with inner ucp coefficients ( y k ) ⊂ M (y_k)\subset M ( y k ) ⊂ M for Q φ Q_\varphi Q φ having the following properties:
(i) The hyperstate is given by φ ( T ) = ∑ k ⟨ T y k ∗ ξ , y k ∗ ξ ⟩ \varphi(T)=\sum_k\langle Ty_k^*\xi,y_k^*\xi\rangle φ ( T ) = ∑ k ⟨ T y k ∗ ξ , y k ∗ ξ ⟩ and satisfies
φ ( J T ∗ J ) = φ ( T ) ( T ∈ B ( L 2 ( M ) ) ) . (12) \varphi(JT^*J)=\varphi(T)\qquad(T\in B(L^2(M))). \tag*{(12)} φ ( J T ∗ J ) = φ ( T ) ( T ∈ B ( L 2 ( M ))) . (12) In particular, the M M M -inner ucp map Q ( a ) = ∑ k y k a y k ∗ Q(a)=\sum_k y_kay_k^* Q ( a ) = ∑ k y k a y k ∗ ( a ∈ M ) (a\in M) ( a ∈ M ) satisfies φ ∘ Q = φ \varphi\circ Q=\varphi φ ∘ Q = φ .
(ii) Each y k ∗ y_k^* y k ∗ admits a bounded modular continuation to i / 2 i/2 i /2 , and the map
y k ⟼ σ i / 2 φ ( y k ∗ ) (13) y_k\longmapsto\sigma^\varphi_{i/2}(y_k^*) \tag*{(13)} y k ⟼ σ i /2 φ ( y k ∗ ) (13) permutes the family ( y k ) (y_k) ( y k ) . Moreover, ∑ k ∥ y k ∥ ∞ 2 < ∞ \sum_k\|y_k\|_\infty^2<\infty ∑ k ∥ y k ∥ ∞ 2 < ∞ .
(iii) If φ \varphi φ is tracial, we may further take y k = λ k u k y_k=\lambda_ku_k y k = λ k u k with u k ∈ U ( M ) u_k\in\mathcal{U}(M) u k ∈ U ( M ) and λ k > 0 \lambda_k>0 λ k > 0 such that ∑ k λ k 2 = 1 \sum_k\lambda_k^2=1 ∑ k λ k 2 = 1 . In particular, Q ( a ) = ∑ k y k a y k ∗ Q(a)=\sum_k y_kay_k^* Q ( a ) = ∑ k y k a y k ∗ is a unitary averaging map.
Proof. Take elements a n ∈ M a_n\in M a n ∈ M , entire analytic for σ φ \sigma^\varphi σ φ , whose linear span is weakly dense in M M M and such that { a n ξ ∣ n ≥ 1 } \{a_n\xi\mid n\geq1\} { a n ξ ∣ n ≥ 1 } has dense linear span in L 2 ( M ) L^2(M) L 2 ( M ) . We may construct such a sequence by applying the Gaussian regularizations
a ⟼ ( r / π ) 1 / 2 ∫ R e − r t 2 σ t φ ( a ) d t ( r ∈ N ) , a\longmapsto(r/\pi)^{1/2}\int_{\mathbb{R}}e^{-rt^2}\sigma_t^\varphi(a)\,dt\qquad(r\in\mathbb{N}), a ⟼ ( r / π ) 1/2 ∫ R e − r t 2 σ t φ ( a ) d t ( r ∈ N ) ,
to a countable strongly dense subset of the unit ball, see [41 Lemma VIII.2.3 and its proof, pp. 98–99] .
Let b n = σ − i / 2 ϕ ( a n ∗ ) b_n=\sigma^\phi_{-i/2}(a_n^*) b n = σ − i /2 ϕ ( a n ∗ ) . By the standard-form identities, we have
J a n J ξ = b n ξ , a n ξ = J b n J ξ . (14) Ja_nJ\xi=b_n\xi,\qquad a_n\xi=Jb_nJ\xi. \tag*{(14)} J a n J ξ = b n ξ , a n ξ = J b n J ξ . (14)
Take c n > 0 c_n>0 c n > 0 such that
∑ n c n ( ∥ a n ∥ ∞ 2 + ∥ b n ∥ ∞ 2 ) < ∞ . (15) \sum_n c_n\left(\lVert a_n\rVert_\infty^2+\lVert b_n\rVert_\infty^2\right)<\infty. \tag*{(15)} n ∑ c n ( ∥ a n ∥ ∞ 2 + ∥ b n ∥ ∞ 2 ) < ∞. (15)
Rescaling the coefficients, we may assume that
C = ∑ n c n ( a n ∗ a n + b n ∗ b n ) , ∥ C ∥ ∞ = 1 , (16) C=\sum_n c_n(a_n^*a_n+b_n^*b_n),\qquad\lVert C\rVert_\infty=1, \tag*{(16)} C = n ∑ c n ( a n ∗ a n + b n ∗ b n ) , ∥ C ∥ ∞ = 1 , (16)
where the series converges in norm. In particular, 0 ≤ C ≤ 1 0\le C\le1 0 ≤ C ≤ 1 . For η ∈ L 2 ( M ) \eta\in L^2(M) η ∈ L 2 ( M ) , write ω η ( T ) = ⟨ T η , η ⟩ \omega_\eta(T)=\langle T\eta,\eta\rangle ω η ( T ) = ⟨ T η , η ⟩ , and set
ψ = ∑ n c n ( ω a n ξ + ω b n ξ ) ∣ M . \psi=\left.\sum_n c_n(\omega_{a_n\xi}+\omega_{b_n\xi})\right|_M. ψ = n ∑ c n ( ω a n ξ + ω b n ξ ) M .
Since ω a n ξ + ω b n ξ = ω J b n J ξ + ω J a n J ξ \omega_{a_n\xi}+\omega_{b_n\xi}=\omega_{Jb_nJ\xi}+\omega_{Ja_nJ\xi} ω a n ξ + ω b n ξ = ω J b n J ξ + ω J a n J ξ by (14) and J a n J , J b n J ∈ M ′ Ja_nJ,Jb_nJ\in M' J a n J , J b n J ∈ M ′ , for x ∈ M + x\in M_+ x ∈ M + we have
ψ ( x ) = ⟨ J C J x 1 / 2 ξ , x 1 / 2 ξ ⟩ ≤ ϕ ( x ) . \psi(x)=\langle JCJx^{1/2}\xi,x^{1/2}\xi\rangle\le\phi(x). ψ ( x ) = ⟨ J C J x 1/2 ξ , x 1/2 ξ ⟩ ≤ ϕ ( x ) .
By [15 Lemma 2.10(1)] , let η ∈ L 2 ( M ) + \eta\in L^2(M)_+ η ∈ L 2 ( M ) + with ϕ − ψ = ⟨ ⋅ η , η ⟩ \phi-\psi=\langle\mathord\cdot\,\eta,\eta\rangle ϕ − ψ = ⟨ ⋅ η , η ⟩ , and define
φ = ∑ n c n ( ω a n ξ + ω b n ξ ) + ω η . \varphi=\sum_n c_n(\omega_{a_n\xi}+\omega_{b_n\xi})+\omega_\eta. φ = n ∑ c n ( ω a n ξ + ω b n ξ ) + ω η .
Then φ ∣ M = ϕ \varphi|_M=\phi φ ∣ M = ϕ . Since J η = η J\eta=\eta J η = η and ω v ( J T ∗ J ) = ω J v ( T ) \omega_v(JT^*J)=\omega_{Jv}(T) ω v ( J T ∗ J ) = ω J v ( T ) , the state φ \varphi φ is symmetric and hence regular. The totality of ( a n ξ ) (a_n\xi) ( a n ξ ) makes φ \varphi φ faithful on B ( L 2 ( M ) ) B(L^2(M)) B ( L 2 ( M )) .
By the same argument as in the proof of Proposition 3.2 , the inequality ϕ − ψ ≤ ϕ \phi-\psi\le\phi ϕ − ψ ≤ ϕ implies that the map x ξ ↦ x η x\xi\mapsto x\eta x ξ ↦ x η extends to a contraction J w J ∈ M ′ JwJ\in M' J w J ∈ M ′ with w ∈ M w\in M w ∈ M . Then
w ξ = η = J w J ξ , ∥ w ∥ ∞ ≤ 1. w\xi=\eta=JwJ\xi,\qquad\lVert w\rVert_\infty\le1. w ξ = η = J w J ξ , ∥ w ∥ ∞ ≤ 1.
The two coefficient families associated with φ \varphi φ are
( z k ) = ( c n b n , c n a n ) n ≥ 1 ∪ { w } , ( y k ) = ( c n a n ∗ , c n b n ∗ ) n ≥ 1 ∪ { w ∗ } . (z_k)=\left(\sqrt{c_n}b_n,\sqrt{c_n}a_n\right)_{n\ge1}\cup\{w\},\qquad
(y_k)=\left(\sqrt{c_n}a_n^*,\sqrt{c_n}b_n^*\right)_{n\ge1}\cup\{w^*\}. ( z k ) = ( c n b n , c n a n ) n ≥ 1 ∪ { w } , ( y k ) = ( c n a n ∗ , c n b n ∗ ) n ≥ 1 ∪ { w ∗ } .
The two representations are
φ ( T ) = ∑ k ⟨ T J z k J ξ , J z k J ξ ⟩ = ∑ k ⟨ T y k ∗ ξ , y k ∗ ξ ⟩ ( T ∈ B ( L 2 ( M ) ) ) . (17) \varphi(T)=\sum_k\langle T Jz_kJ\xi,Jz_kJ\xi\rangle
=\sum_k\langle T y_k^*\xi,y_k^*\xi\rangle
\qquad(T\in B(L^2(M))). \tag*{(17)} φ ( T ) = k ∑ ⟨ T J z k J ξ , J z k J ξ ⟩ = k ∑ ⟨ T y k ∗ ξ , y k ∗ ξ ⟩ ( T ∈ B ( L 2 ( M ))) . (17)
Since φ ∣ M = ϕ \varphi|_M=\phi φ ∣ M = ϕ and φ ( J a ∗ J ) = ϕ ( a ) \varphi(Ja^*J)=\phi(a) φ ( J a ∗ J ) = ϕ ( a ) for a ∈ M a\in M a ∈ M ,
∑ k z k ∗ z k = 1 , ∑ k y k y k ∗ = 1. \sum_k z_k^*z_k=1,\qquad\sum_k y_ky_k^*=1. k ∑ z k ∗ z k = 1 , k ∑ y k y k ∗ = 1.
Moreover, since a n ∈ C [ ( z k ) ] ‾ a_n\in\overline{\mathop{\mathrm{C}}[(z_k)]} a n ∈ C [( z k )] and a n ∗ ∈ C [ ( y k ) ] ‾ a_n^*\in\overline{\mathop{\mathrm{C}}[(y_k)]} a n ∗ ∈ C [( y k )] , we have C [ ( z k ) ] ‾ w = C [ ( y k ) ] ‾ w = M \overline{\mathop{\mathrm{C}}[(z_k)]}^{\,w}=\overline{\mathop{\mathrm{C}}[(y_k)]}^{\,w}=M C [( z k )] w = C [( y k )] w = M , so φ \varphi φ is nrsg. By (17) ,
ϕ ( Q ( a ) ) = ∑ k ⟨ a y k ∗ ξ , y k ∗ ξ ⟩ = φ ( a ) = ϕ ( a ) ( a ∈ M ) . \phi(Q(a))=\sum_k\langle ay_k^*\xi,y_k^*\xi\rangle=\varphi(a)=\phi(a)\qquad(a\in M). ϕ ( Q ( a )) = k ∑ ⟨ a y k ∗ ξ , y k ∗ ξ ⟩ = φ ( a ) = ϕ ( a ) ( a ∈ M ) .
This proves (i).
To prove (ii), note that
σ i / 2 ϕ ( a n ) = b n ∗ , σ i / 2 ϕ ( b n ) = a n ∗ . \sigma^\phi_{i/2}(a_n)=b_n^*,\qquad\sigma^\phi_{i/2}(b_n)=a_n^*. σ i /2 ϕ ( a n ) = b n ∗ , σ i /2 ϕ ( b n ) = a n ∗ .
It remains to treat w ∗ w^* w ∗ . For r > 0 r>0 r > 0 , define
w r = ( r / π ) 1 / 2 ∫ R e − r t 2 σ t ϕ ( w ) d t . w_r=(r/\pi)^{1/2}\int_{\mathbb R}e^{-rt^2}\sigma_t^\phi(w)\,dt. w r = ( r / π ) 1/2 ∫ R e − r t 2 σ t ϕ ( w ) d t .
By the identity J w ξ = w ∗ ξ Jw\xi=w^{*}\xi J w ξ = w ∗ ξ and the commutation relation J Δ ϕ i t = Δ ϕ i t J J\Delta_{\phi}^{it}=\Delta_{\phi}^{it}J J Δ ϕ i t = Δ ϕ i t J , we have J w r ξ = w r ξ Jw_{r}\xi=w_{r}\xi J w r ξ = w r ξ . The Tomita identity and the separating property of ξ \xi ξ imply
σ − i / 2 ϕ ( w r ∗ ) = w r , σ i / 2 ϕ ( w r ) = w r ∗ . \sigma_{-i/2}^{\phi}(w_{r}^{*})=w_{r},\qquad\sigma_{i/2}^{\phi}(w_{r})=w_{r}^{*}. σ − i /2 ϕ ( w r ∗ ) = w r , σ i /2 ϕ ( w r ) = w r ∗ .
The boundary values on the half-strip are bounded by ∥ w ∥ ∞ \lVert w\rVert_{\infty} ∥ w ∥ ∞ . By the three-lines theorem and the ultraweak closedness of the analytic generator, we have σ i / 2 ϕ ( w ) = w ∗ \sigma_{i/2}^{\phi}(w)=w^{*} σ i /2 ϕ ( w ) = w ∗ , see [18 p. 192] . Hence (13) exchanges the paired coefficients and fixes w ∗ w^{*} w ∗ .
Moreover, by the choice of c n c_{n} c n in (15) and the definition of ( y k ) (y_{k}) ( y k ) , we have
∑ k ∥ y k ∥ ∞ 2 = ∑ n c n ( ∥ a n ∥ ∞ 2 + ∥ b n ∥ ∞ 2 ) + ∥ w ∥ ∞ 2 < ∞ . \sum_{k}\lVert y_{k}\rVert_{\infty}^{2}=\sum_{n}c_{n}\left(\lVert a_{n}\rVert_{\infty}^{2}+\lVert b_{n}\rVert_{\infty}^{2}\right)+\lVert w\rVert_{\infty}^{2}<\infty. k ∑ ∥ y k ∥ ∞ 2 = n ∑ c n ( ∥ a n ∥ ∞ 2 + ∥ b n ∥ ∞ 2 ) + ∥ w ∥ ∞ 2 < ∞.
This proves (ii).
Proposition 3.10 (cf. [9 Propositions 2.5(i) and 2.9] ). For k ≥ 0 k\geq0 k ≥ 0 , define the k k k -th convolution φ ∗ k \varphi^{*k} φ ∗ k by φ ∗ k ( T ) = ⟨ P φ k ( T ) ξ φ , ξ φ ⟩ \varphi^{*k}(T)=\langle P_{\varphi}^{k}(T)\xi_{\varphi},\xi_{\varphi}\rangle φ ∗ k ( T ) = ⟨ P φ k ( T ) ξ φ , ξ φ ⟩ . Then the normal state φ ‾ = ∑ n = 0 ∞ 2 − n − 1 φ ∗ n \overline{\varphi}=\sum_{n=0}^{\infty}2^{-n-1}\varphi^{*n} φ = ∑ n = 0 ∞ 2 − n − 1 φ ∗ n is faithful on B ( L 2 ( M ) ) B(L^{2}(M)) B ( L 2 ( M )) . In particular, ζ = φ ∣ Har ( P φ ) \zeta=\varphi|_{\operatorname{Har}(P_{\varphi})} ζ = φ ∣ Har ( P φ ) is a normal faithful state on the Poisson boundary B φ B_{\varphi} B φ . Moreover, M M M is the largest von Neumann subalgebra of Har ( P φ ) \operatorname{Har}(P_{\varphi}) Har ( P φ ) for the original operator product on B ( L 2 ( M ) ) B(L^{2}(M)) B ( L 2 ( M )) .
Proof. Let T ∈ B ( L 2 ( M ) ) + T\in B(L^{2}(M))_{+} T ∈ B ( L 2 ( M ) ) + satisfy φ ‾ ( T ) = 0 \overline{\varphi}(T)=0 φ ( T ) = 0 . Since all the coefficients are positive, φ ∗ n ( T ) = 0 \varphi^{*n}(T)=0 φ ∗ n ( T ) = 0 for every n ≥ 0 n\geq0 n ≥ 0 . Expanding the iterates of P φ P_{\varphi} P φ and using positivity, we have
⟨ T ( J w J ) ξ φ , ( J w J ) ξ φ ⟩ = 0 \langle T(JwJ)\xi_{\varphi},(JwJ)\xi_{\varphi}\rangle=0 ⟨ T ( J w J ) ξ φ , ( J w J ) ξ φ ⟩ = 0
for every word w w w in ( z k ) (z_{k}) ( z k ) , including the empty word. The weak density of C [ ( z k ) ] C[(z_{k})] C [( z k )] implies that the vectors ( J w J ) ξ φ (JwJ)\xi_{\varphi} ( J w J ) ξ φ have dense linear span in L 2 ( M ) L^{2}(M) L 2 ( M ) . Hence T = 0 T=0 T = 0 , proving that φ ‾ \overline{\varphi} φ is faithful. Since φ ∗ n ( T ) = φ ( T ) \varphi^{*n}(T)=\varphi(T) φ ∗ n ( T ) = φ ( T ) for T ∈ Har ( P φ ) T\in\operatorname{Har}(P_{\varphi}) T ∈ Har ( P φ ) and n ≥ 0 n\geq0 n ≥ 0 , we have ζ = φ ‾ ∣ Har ( P φ ) \zeta=\overline{\varphi}|_{\operatorname{Har}(P_{\varphi})} ζ = φ ∣ Har ( P φ ) . Hence ζ \zeta ζ is normal and faithful on B φ B_{\varphi} B φ .
Let N ⊂ Har ( P φ ) \mathcal{N}\subset\operatorname{Har}(P_{\varphi}) N ⊂ Har ( P φ ) be a von Neumann subalgebra for the original operator product, and let v k = J z k ∗ J v_{k}=Jz_{k}^{*}J v k = J z k ∗ J . For T ∈ N T\in\mathcal{N} T ∈ N , the operators T , T ∗ , T T ∗ T,T^{*},TT^{*} T , T ∗ , T T ∗ are fixed by P φ P_{\varphi} P φ , so
∑ k ( v k T − T v k ) ( v k T − T v k ) ∗ = P φ ( T T ∗ ) − P φ ( T ) T ∗ − T P φ ( T ∗ ) + T T ∗ = 0. (18) \sum_{k}(v_{k}T-Tv_{k})(v_{k}T-Tv_{k})^{*}
=P_{\varphi}(TT^{*})-P_{\varphi}(T)T^{*}-TP_{\varphi}(T^{*})+TT^{*}=0. \tag*{(18)} k ∑ ( v k T − T v k ) ( v k T − T v k ) ∗ = P φ ( T T ∗ ) − P φ ( T ) T ∗ − T P φ ( T ∗ ) + T T ∗ = 0. (18)
Hence [ T , v k ] = 0 [T,v_{k}]=0 [ T , v k ] = 0 for every k k k . By applying (18) to T ∗ T^{*} T ∗ , we also have [ T , v k ∗ ] = 0 [T,v_{k}^{*}]=0 [ T , v k ∗ ] = 0 for every k k k . By the weak density of C [ ( z k ) ] C[(z_{k})] C [( z k )] , we have T ∈ ( J M J ) ′ = M T\in(JMJ)'=M T ∈ ( J M J ) ′ = M . Since M M M itself is contained in the harmonic space, it is the largest such subalgebra.
The following is a well-known result [9 Lemma 2.6] , which also appears in earlier literature such as [10 ] , [3 Lemma 3.4] , and [4 Lemma 3.1] . For completeness, we include the proof.
Lemma 3.11. Let A \mathcal{A} A be a von Neumann algebra, and let P : A → A \mathcal{P}:\mathcal{A}\to\mathcal{A} P : A → A be a normal ucp map preserving a normal faithful state ψ A \psi_{\mathcal{A}} ψ A on A \mathcal{A} A . Then Har ( P , A ) = { x ∈ A ∣ P ( x ) = x } \operatorname{Har}(\mathcal{P},\mathcal{A})=\{x\in\mathcal{A}\mid\mathcal{P}(x)=x\} Har ( P , A ) = { x ∈ A ∣ P ( x ) = x } is a von Neumann subalgebra of A \mathcal{A} A .
Proof. Since Har ( P , A ) \operatorname{Har}(\mathcal{P},\mathcal{A}) Har ( P , A ) is a weakly closed operator system, by polarization we only need to prove that a a ∗ ∈ Har ( P , A ) aa^{*} \in\operatorname{Har}(\mathcal{P},\mathcal{A}) a a ∗ ∈ Har ( P , A ) for every a ∈ Har ( P , A ) a \in\operatorname{Har}(\mathcal{P},\mathcal{A}) a ∈ Har ( P , A ) . For such an a a a , Kadison’s inequality gives P ( a a ∗ ) − a a ∗ = P ( a a ∗ ) − P ( a ) P ( a ) ∗ ≥ 0 \mathcal{P}(aa^{*})-aa^{*}=\mathcal{P}(aa^{*})-\mathcal{P}(a)\mathcal{P}(a)^{*}\geq0 P ( a a ∗ ) − a a ∗ = P ( a a ∗ ) − P ( a ) P ( a ) ∗ ≥ 0 . Since ψ A ∘ P = ψ A \psi_{\mathcal{A}}\circ\mathcal{P}=\psi_{\mathcal{A}} ψ A ∘ P = ψ A , we have ψ A ( P ( a a ∗ ) − a a ∗ ) = 0 \psi_{\mathcal{A}}(\mathcal{P}(aa^{*})-aa^{*})=0 ψ A ( P ( a a ∗ ) − a a ∗ ) = 0 . By faithfulness, P ( a a ∗ ) = a a ∗ \mathcal{P}(aa^{*})=aa^{*} P ( a a ∗ ) = a a ∗ , which finishes the proof. □
Proposition 3.12 (cf. [9 Proposition 2.7] ). We have the ergodicity of the Poisson boundary , that is, M ′ ∩ B φ = Z ( M ) M^{\prime}\cap B_{\varphi}=\mathcal{Z}(M) M ′ ∩ B φ = Z ( M ) .
Proof. By identifying B φ B_{\varphi} B φ with Har ( P φ ) \operatorname{Har}(\mathcal{P}_{\varphi}) Har ( P φ ) , we consider elements in M ′ ∩ Har ( P φ ) M^{\prime}\cap\operatorname{Har}(\mathcal{P}_{\varphi}) M ′ ∩ Har ( P φ ) . Note that M ′ ∩ B ( L 2 ( M ) ) = J M J M^{\prime}\cap B(L^{2}(M))=JMJ M ′ ∩ B ( L 2 ( M )) = J M J . Hence, we only need to compute J M J ∩ Har ( P φ ) JMJ\cap\operatorname{Har}(\mathcal{P}_{\varphi}) J M J ∩ Har ( P φ ) .
Write φ o = φ ∣ J M J \varphi^{\mathrm{o}}=\varphi|_{JMJ} φ o = φ ∣ J M J . Since P φ \mathcal{P}_{\varphi} P φ is J M J JMJ J M J -inner, we have P φ ( J M J ) ⊂ J M J \mathcal{P}_{\varphi}(JMJ)\subset JMJ P φ ( J M J ) ⊂ J M J . Since φ \varphi φ is regular, for any T ∈ J M J T\in JMJ T ∈ J M J , we have
φ o ( P φ ( T ) ) = ⟨ P φ ( T ) ξ φ , ξ φ ⟩ = φ ( T ) = φ o ( T ) . \varphi^{\mathrm{o}}(\mathcal{P}_{\varphi}(T))=\langle\mathcal{P}_{\varphi}(T)\xi_{\varphi},\xi_{\varphi}\rangle=\varphi(T)=\varphi^{\mathrm{o}}(T). φ o ( P φ ( T )) = ⟨ P φ ( T ) ξ φ , ξ φ ⟩ = φ ( T ) = φ o ( T ) .
Here the first equality uses P φ ( T ) ∈ J M J \mathcal{P}_{\varphi}(T)\in JMJ P φ ( T ) ∈ J M J and the fact that ξ φ \xi_{\varphi} ξ φ implements φ o \varphi^{\mathrm{o}} φ o on J M J JMJ J M J . Hence, the restriction P φ ∣ J M J \mathcal{P}_{\varphi}|_{JMJ} P φ ∣ J M J preserves the normal faithful state φ o \varphi^{\mathrm{o}} φ o on J M J JMJ J M J . By Lemma 3.11, the fixed-point space Har ( P φ ∣ J M J , J M J ) \operatorname{Har}(\mathcal{P}_{\varphi}|_{JMJ},JMJ) Har ( P φ ∣ J M J , J M J ) is a von Neumann subalgebra of J M J JMJ J M J .
Because φ \varphi φ is strongly generating, by Proposition 3.10, M M M is the largest von Neumann subalgebra contained in Har ( P φ ) \operatorname{Har}(\mathcal{P}_{\varphi}) Har ( P φ ) . Therefore, Har ( P φ ) ∩ J M J ⊂ M \operatorname{Har}(\mathcal{P}_{\varphi})\cap JMJ\subset M Har ( P φ ) ∩ J M J ⊂ M , which implies Har ( P φ ) ∩ J M J ⊂ M ∩ J M J = Z ( M ) \operatorname{Har}(\mathcal{P}_{\varphi})\cap JMJ\subset M\cap JMJ=\mathcal{Z}(M) Har ( P φ ) ∩ J M J ⊂ M ∩ J M J = Z ( M ) . Since Z ( M ) ⊂ Har ( P φ ) \mathcal{Z}(M)\subset\operatorname{Har}(\mathcal{P}_{\varphi}) Z ( M ) ⊂ Har ( P φ ) is trivial, we conclude M ′ ∩ B φ = Z ( M ) M^{\prime}\cap B_{\varphi}=\mathcal{Z}(M) M ′ ∩ B φ = Z ( M ) . □
We next prove the double ergodicity of the Poisson boundary.
Proposition 3.13. Let M ⊂ A M\subset\mathcal{A} M ⊂ A be a W ∗ W^{*} W ∗ -inclusion. Using the coefficients ( y k ) (y_{k}) ( y k ) of Q φ Q_{\varphi} Q φ as in (1) , we define the normal M M M -inner ucp map Q : A → A Q:\mathcal{A}\to\mathcal{A} Q : A → A by Q ( T ) = ∑ k = 1 ∞ y k T y k ∗ Q(T)=\sum_{k=1}^{\infty}y_{k}Ty_{k}^{*} Q ( T ) = ∑ k = 1 ∞ y k T y k ∗ . Let
Har ( Q , A ) = { T ∈ A ∣ Q ( T ) = T } . \operatorname{Har}(Q,\mathcal{A})=\{T\in\mathcal{A}\mid Q(T)=T\}. Har ( Q , A ) = { T ∈ A ∣ Q ( T ) = T } .
Assume there exists a normal faithful ucp map Φ : A → B φ \Phi:\mathcal{A}\to B_{\varphi} Φ : A → B φ such that Φ ∣ M = id M \Phi|_{M}=\operatorname{id}_{M} Φ ∣ M = id M . Then we have
Har ( Q , A ) = M ′ ∩ A . \operatorname{Har}(Q,\mathcal{A})=M^{\prime}\cap\mathcal{A}. Har ( Q , A ) = M ′ ∩ A .
In particular , we have the double ergodicity of the Poisson boundary (cf. [9 Theorem 3.1] )
Har ( P φ ) ∩ Har ( Q φ ) = Z ( M ) . \operatorname{Har}(\mathcal{P}_{\varphi})\cap\operatorname{Har}(Q_{\varphi})=\mathcal{Z}(M). Har ( P φ ) ∩ Har ( Q φ ) = Z ( M ) .
Proof. Let ζ \zeta ζ be the normal faithful stationary state on the Poisson boundary B φ B_{\varphi} B φ . We define a normal faithful state ψ A \psi_{\mathcal{A}} ψ A on A \mathcal{A} A by ψ A = ζ ∘ Φ \psi_{\mathcal{A}}=\zeta\circ\Phi ψ A = ζ ∘ Φ .
Because Φ ∣ M = id M \Phi|_{M}=\operatorname{id}_{M} Φ ∣ M = id M , Φ \Phi Φ is an M M M -bimodular map. Hence Φ ∘ Q = Q φ ∘ Φ \Phi\circ Q=Q_{\varphi}\circ\Phi Φ ∘ Q = Q φ ∘ Φ . Indeed, for every T ∈ A T\in\mathcal{A} T ∈ A ,
Φ ( Q ( T ) ) = ∑ k = 1 ∞ Φ ( y k T y k ∗ ) = ∑ k = 1 ∞ y k Φ ( T ) y k ∗ = Q φ ( Φ ( T ) ) . (19) \Phi(Q(T))=\sum_{k=1}^{\infty}\Phi(y_{k}Ty_{k}^{*})=\sum_{k=1}^{\infty}y_{k}\Phi(T)y_{k}^{*}=Q_{\varphi}(\Phi(T)). \tag*{(19)} Φ ( Q ( T )) = k = 1 ∑ ∞ Φ ( y k T y k ∗ ) = k = 1 ∑ ∞ y k Φ ( T ) y k ∗ = Q φ ( Φ ( T )) . (19)
We now prove that ζ \zeta ζ is Q φ Q_{\varphi} Q φ -invariant on B φ B_{\varphi} B φ . Identifying B φ B_{\varphi} B φ with Har ( P φ ) \operatorname{Har}(\mathcal{P}_{\varphi}) Har ( P φ ) , the state is implemented by the vector state ζ ( ⋅ ) = ⟨ P φ ( ⋅ ) ξ φ , ξ φ ⟩ = ⟨ ⋅ ξ φ , ξ φ ⟩ \zeta(\mathord{\cdot})=\langle\mathcal{P}_{\varphi}(\mathord{\cdot})\xi_{\varphi},\xi_{\varphi}\rangle=\langle\mathord{\cdot}\xi_{\varphi},\xi_{\varphi}\rangle ζ ( ⋅ ) = ⟨ P φ ( ⋅ ) ξ φ , ξ φ ⟩ = ⟨ ⋅ ξ φ , ξ φ ⟩ , which coincides with the original hyperstate φ \varphi φ on this space. By the defining construction of Q φ Q_{\varphi} Q φ , we have ⟨ Q φ ( T ) ξ φ , ξ φ ⟩ = φ ( T ) \langle Q_{\varphi}(T)\xi_{\varphi},\xi_{\varphi}\rangle=\varphi(T) ⟨ Q φ ( T ) ξ φ , ξ φ ⟩ = φ ( T ) for any T ∈ B ( L 2 ( M ) ) T\in B(L^{2}(M)) T ∈ B ( L 2 ( M )) . Hence, for any S ∈ B φ S\in B_{\varphi} S ∈ B φ , we have
ζ ( Q φ ( S ) ) = ⟨ Q φ ( S ) ξ φ , ξ φ ⟩ = φ ( S ) = ζ ( S ) . (20) \zeta(Q_{\varphi}(S))=\langle Q_{\varphi}(S)\xi_{\varphi},\xi_{\varphi}\rangle=\varphi(S)=\zeta(S). \tag*{(20)} ζ ( Q φ ( S )) = ⟨ Q φ ( S ) ξ φ , ξ φ ⟩ = φ ( S ) = ζ ( S ) . (20)
By (19) and (20) , for T ∈ A T\in\mathcal{A} T ∈ A ,
ψ A ( Q ( T ) ) = ζ ( Φ ( Q ( T ) ) ) = ζ ( Q φ ( Φ ( T ) ) ) = ζ ( Φ ( T ) ) = ψ A ( T ) . \psi_{\mathcal{A}}(Q(T))=\zeta(\Phi(Q(T)))=\zeta(Q_{\varphi}(\Phi(T)))=\zeta(\Phi(T))=\psi_{\mathcal{A}}(T). ψ A ( Q ( T )) = ζ ( Φ ( Q ( T ))) = ζ ( Q φ ( Φ ( T ))) = ζ ( Φ ( T )) = ψ A ( T ) .
This shows that Q Q Q preserves the normal faithful state ψ A \psi_{\mathcal{A}} ψ A . By Lemma 3.11, the fixed-point space A 0 = Har ( Q , A ) \mathcal{A}_{0}=\operatorname{Har}(Q,\mathcal{A}) A 0 = Har ( Q , A ) is a von Neumann subalgebra of A \mathcal{A} A .
Clearly, if T ∈ M ′ ∩ A T\in M^{\prime}\cap\mathcal{A} T ∈ M ′ ∩ A , then Q ( T ) = ∑ k = 1 ∞ y k T y k ∗ = ∑ k = 1 ∞ y k y k ∗ T = T Q(T)=\sum_{k=1}^{\infty}y_{k}Ty_{k}^{*}=\sum_{k=1}^{\infty}y_{k}y_{k}^{*}T=T Q ( T ) = ∑ k = 1 ∞ y k T y k ∗ = ∑ k = 1 ∞ y k y k ∗ T = T , so M ′ ∩ A ⊂ A 0 M^{\prime}\cap\mathcal{A}\subset\mathcal{A}_{0} M ′ ∩ A ⊂ A 0 . Conversely, suppose T ∈ A 0 T \in\mathcal{A}_0 T ∈ A 0 . Because A 0 \mathcal{A}_0 A 0 is a von Neumann subalgebra, we also have T T ∗ ∈ A 0 TT^* \in\mathcal{A}_0 T T ∗ ∈ A 0 , meaning Q ( T ) = T \mathcal{Q}(T)=T Q ( T ) = T and Q ( T T ∗ ) = T T ∗ \mathcal{Q}(TT^*)=TT^* Q ( T T ∗ ) = T T ∗ . We compute the following sum of positive elements
∑ k = 1 ∞ ( y k T − T y k ) ( y k T − T y k ) ∗ = ∑ k = 1 ∞ ( y k T T ∗ y k ∗ − y k T y k ∗ T ∗ − T y k T ∗ y k ∗ + T y k y k ∗ T ∗ ) = Q ( T T ∗ ) − Q ( T ) T ∗ − T Q ( T ∗ ) + T ( ∑ k = 1 ∞ y k y k ∗ ) T ∗ = T T ∗ − T T ∗ − T T ∗ + T T ∗ = 0. (21) \begin{aligned}
\sum_{k=1}^{\infty}(y_kT-Ty_k)(y_kT-Ty_k)^*
&=\sum_{k=1}^{\infty}(y_kTT^*y_k^*-y_kTy_k^*T^*-Ty_kT^*y_k^*+Ty_ky_k^*T^*)\\
&=\mathcal{Q}(TT^*)-\mathcal{Q}(T)T^*-T\mathcal{Q}(T^*)+T\left(\sum_{k=1}^{\infty}y_ky_k^*\right)T^*\\
&=TT^*-TT^*-TT^*+TT^*=0. \tag*{(21)}
\end{aligned} k = 1 ∑ ∞ ( y k T − T y k ) ( y k T − T y k ) ∗ = k = 1 ∑ ∞ ( y k T T ∗ y k ∗ − y k T y k ∗ T ∗ − T y k T ∗ y k ∗ + T y k y k ∗ T ∗ ) = Q ( T T ∗ ) − Q ( T ) T ∗ − T Q ( T ∗ ) + T ( k = 1 ∑ ∞ y k y k ∗ ) T ∗ = T T ∗ − T T ∗ − T T ∗ + T T ∗ = 0. (21)
By (21) , [ y k , T ] = 0 [y_k,T]=0 [ y k , T ] = 0 for every k k k . Since C [ ( y k ) ] ‾ w = M \overline{\mathbb{C}[(y_k)]}^{w}=M C [( y k )] w = M , T ∈ M ′ ∩ A T\in M'\cap\mathcal{A} T ∈ M ′ ∩ A . Hence Har ( Q , A ) = M ′ ∩ A \operatorname{Har}(\mathcal{Q},\mathcal{A})=M'\cap\mathcal{A} Har ( Q , A ) = M ′ ∩ A .
For the double ergodicity of B φ \mathcal{B}_{\varphi} B φ , we set A = B φ ≅ Har ( P φ ) \mathcal{A}=\mathcal{B}_{\varphi}\cong\operatorname{Har}(\mathcal{P}_{\varphi}) A = B φ ≅ Har ( P φ ) and let Φ = id B φ \Phi=\operatorname{id}_{\mathcal{B}_{\varphi}} Φ = id B φ . The hypothesis is satisfied, so we have
Har ( P φ ) ∩ Har ( Q φ ) = Har ( Q , B φ ) = M ′ ∩ B φ . \operatorname{Har}(\mathcal{P}_{\varphi})\cap\operatorname{Har}(\mathcal{Q}_{\varphi})=\operatorname{Har}(\mathcal{Q},\mathcal{B}_{\varphi})=M'\cap\mathcal{B}_{\varphi}. Har ( P φ ) ∩ Har ( Q φ ) = Har ( Q , B φ ) = M ′ ∩ B φ .
By Proposition 3.12, M ′ ∩ B φ = Z ( M ) M'\cap\mathcal{B}_{\varphi}=\mathcal{Z}(M) M ′ ∩ B φ = Z ( M ) , which concludes the proof. □
By taking the Cesàro means of Q \mathcal{Q} Q , we have the inner UCP-Dixmier property.
Corollary 3.14. Let M ⊂ A M\subset\mathcal{A} M ⊂ A be a W ∗ W^* W ∗ -inclusion, and assume there exists a normal faithful ucp map Φ : A → B φ \Phi:\mathcal{A}\to\mathcal{B}_{\varphi} Φ : A → B φ such that Φ ∣ M = id M \Phi|_M=\operatorname{id}_M Φ ∣ M = id M . Then M ⊂ A M\subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property.
Proof. Let Q \mathcal{Q} Q be as in Proposition 3.13, and let C n = ( 1 / n ) ∑ k = 0 n − 1 Q k C_n=(1/n)\sum_{k=0}^{n-1}\mathcal{Q}^k C n = ( 1/ n ) ∑ k = 0 n − 1 Q k . Let ω ∈ β N ∖ N \omega\in\beta\mathbb{N}\setminus\mathbb{N} ω ∈ β N ∖ N be a free ultrafilter. By Proposition 3.13, we can define a conditional expectation E : A → M ′ ∩ A E:\mathcal{A}\to M'\cap\mathcal{A} E : A → M ′ ∩ A by
E ( x ) : = w o - l i m n → ω C n ( x ) . E(x):=\mathop{\mathrm{wo\text{-}lim}}_{n\to\omega}C_n(x). E ( x ) := wo - lim n → ω C n ( x ) . Recall that Q \mathcal{Q} Q is an M M M -inner ucp map. Therefore, E ( x ) E(x) E ( x ) lies in the weak closure of the inner UCP orbit of x x x :
E ( x ) ∈ { P ( x ) ∣ P is an M -inner ucp map } ‾ w ∩ ( M ′ ∩ A ) . E(x)\in\overline{\{\mathcal{P}(x)\mid\mathcal{P}\text{ is an }M\text{-inner ucp map}\}}^{w}\cap(M'\cap\mathcal{A}). E ( x ) ∈ { P ( x ) ∣ P is an M -inner ucp map } w ∩ ( M ′ ∩ A ) . Hence M ⊂ A M\subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property. □
We now prove the rigidity of the Poisson boundary, one of the most important properties of Poisson boundaries in both the group setting and the Das–Peterson framework. Although this result is not used in the subsequent proofs, we include it for completeness of our treatment of Poisson boundary theory.
Proposition 3.15 (cf. [9 Theorem 4.1] ). Let C C C be a weakly closed operator system such that M ⊂ C ⊂ B φ M\subset C\subset\mathcal{B}_{\varphi} M ⊂ C ⊂ B φ and M C + C M ⊂ C MC+CM\subset C M C + C M ⊂ C . If Ψ : C → B φ \Psi:C\to\mathcal{B}_{\varphi} Ψ : C → B φ is a normal ucp map such that Ψ ∣ M = id M \Psi|_M=\operatorname{id}_M Ψ ∣ M = id M , then Ψ = id C \Psi=\operatorname{id}_C Ψ = id C .
Proof. We identify C C C with its image under the isomorphism B φ ≅ Har ( P φ ) \mathcal{B}_{\varphi}\cong\operatorname{Har}(\mathcal{P}_{\varphi}) B φ ≅ Har ( P φ ) . Then Ψ \Psi Ψ is a normal ucp map from C C C to Har ( P φ ) \operatorname{Har}(\mathcal{P}_{\varphi}) Har ( P φ ) . Since Ψ ∣ M = id M \Psi|_M=\operatorname{id}_M Ψ ∣ M = id M , the map Ψ \Psi Ψ is M M M -bimodular. Using the explicit formula Q φ ( T ) = ∑ k = 1 ∞ y k T y k ∗ \mathcal{Q}_{\varphi}(T)=\sum_{k=1}^{\infty}y_kTy_k^* Q φ ( T ) = ∑ k = 1 ∞ y k T y k ∗ , for any T ∈ C T\in C T ∈ C we have
⟨ Ψ ( Q φ ( T ) ) ξ φ , ξ φ ⟩ = ∑ k = 1 ∞ ⟨ Ψ ( y k T y k ∗ ) ξ φ , ξ φ ⟩ = ∑ k = 1 ∞ ⟨ y k Ψ ( T ) y k ∗ ξ φ , ξ φ ⟩ = ⟨ Q φ ( Ψ ( T ) ) ξ φ , ξ φ ⟩ . \langle\Psi(\mathcal{Q}_{\varphi}(T))\xi_{\varphi},\xi_{\varphi}\rangle
=\sum_{k=1}^{\infty}\langle\Psi(y_kTy_k^*)\xi_{\varphi},\xi_{\varphi}\rangle
=\sum_{k=1}^{\infty}\langle y_k\Psi(T)y_k^*\xi_{\varphi},\xi_{\varphi}\rangle
=\langle\mathcal{Q}_{\varphi}(\Psi(T))\xi_{\varphi},\xi_{\varphi}\rangle. ⟨ Ψ ( Q φ ( T )) ξ φ , ξ φ ⟩ = k = 1 ∑ ∞ ⟨ Ψ ( y k T y k ∗ ) ξ φ , ξ φ ⟩ = k = 1 ∑ ∞ ⟨ y k Ψ ( T ) y k ∗ ξ φ , ξ φ ⟩ = ⟨ Q φ ( Ψ ( T )) ξ φ , ξ φ ⟩ . Since Ψ ( T ) ∈ Har ( P φ ) \Psi(T)\in\operatorname{Har}(\mathcal{P}_{\varphi}) Ψ ( T ) ∈ Har ( P φ ) and φ \varphi φ is regular, the term on the right evaluates to
⟨ P φ ( Ψ ( T ) ) ξ φ , ξ φ ⟩ = ⟨ Ψ ( T ) ξ φ , ξ φ ⟩ . \langle\mathcal{P}_{\varphi}(\Psi(T))\xi_{\varphi},\xi_{\varphi}\rangle=\langle\Psi(T)\xi_{\varphi},\xi_{\varphi}\rangle. ⟨ P φ ( Ψ ( T )) ξ φ , ξ φ ⟩ = ⟨ Ψ ( T ) ξ φ , ξ φ ⟩ . Hence, ⟨ Ψ ( Q φ ( T ) ) ξ φ , ξ φ ⟩ = ⟨ Ψ ( T ) ξ φ , ξ φ ⟩ \langle\Psi(\mathcal{Q}_{\varphi}(T))\xi_{\varphi},\xi_{\varphi}\rangle=\langle\Psi(T)\xi_{\varphi},\xi_{\varphi}\rangle ⟨ Ψ ( Q φ ( T )) ξ φ , ξ φ ⟩ = ⟨ Ψ ( T ) ξ φ , ξ φ ⟩ for all T ∈ C T\in C T ∈ C . For the Cesàro means C n = ( 1 / n ) ∑ k = 0 n − 1 Q φ k C_n=(1/n)\sum_{k=0}^{n-1}\mathcal{Q}_{\varphi}^k C n = ( 1/ n ) ∑ k = 0 n − 1 Q φ k , linearity gives
⟨ Ψ ( C n ( T ) ) ξ φ , ξ φ ⟩ = ⟨ Ψ ( T ) ξ φ , ξ φ ⟩ . \langle\Psi(C_n(T))\xi_{\varphi},\xi_{\varphi}\rangle=\langle\Psi(T)\xi_{\varphi},\xi_{\varphi}\rangle. ⟨ Ψ ( C n ( T )) ξ φ , ξ φ ⟩ = ⟨ Ψ ( T ) ξ φ , ξ φ ⟩ .
Let ω ∈ β N ∖ N \omega\in\beta\mathbb{N} \setminus\mathbb{N} ω ∈ β N ∖ N be a free ultrafilter. By normality and Proposition 3.13, we have
z : = wo-lim n → ω C n ( T ) ∈ Har ( P φ ) ∩ Har ( Q φ ) = Z ( M ) . z := \operatorname{wo-lim}_{n \to\omega} C_n(T) \in\operatorname{Har}(\mathcal{P}_{\varphi}) \cap\operatorname{Har}(\mathcal{Q}_{\varphi}) = \mathcal{Z}(M). z := wo-lim n → ω C n ( T ) ∈ Har ( P φ ) ∩ Har ( Q φ ) = Z ( M ) .
Since z ∈ M z \in M z ∈ M , we have Ψ ( z ) = z \Psi(z)=z Ψ ( z ) = z . By normality of Ψ \Psi Ψ , we have the first equality below. The last equality follows from (20) and normality of ζ \zeta ζ :
⟨ Ψ ( T ) ξ φ , ξ φ ⟩ = ⟨ Ψ ( z ) ξ φ , ξ φ ⟩ = ⟨ z ξ φ , ξ φ ⟩ = ⟨ T ξ φ , ξ φ ⟩ . \langle\Psi(T)\xi_{\varphi},\xi_{\varphi}\rangle= \langle\Psi(z)\xi_{\varphi},\xi_{\varphi}\rangle= \langle z\xi_{\varphi},\xi_{\varphi}\rangle= \langle T\xi_{\varphi},\xi_{\varphi}\rangle. ⟨ Ψ ( T ) ξ φ , ξ φ ⟩ = ⟨ Ψ ( z ) ξ φ , ξ φ ⟩ = ⟨ z ξ φ , ξ φ ⟩ = ⟨ T ξ φ , ξ φ ⟩ .
Now, for any a , b ∈ M a,b \in M a , b ∈ M and T ∈ C T \in\mathcal{C} T ∈ C , the element b ∗ T a b^{*}Ta b ∗ T a is still in C \mathcal{C} C . Using the M M M -bimodularity of Ψ \Psi Ψ , we have
⟨ Ψ ( T ) a ξ φ , b ξ φ ⟩ = ⟨ Ψ ( b ∗ T a ) ξ φ , ξ φ ⟩ = ⟨ b ∗ T a ξ φ , ξ φ ⟩ = ⟨ T a ξ φ , b ξ φ ⟩ . \langle\Psi(T)a\xi_{\varphi},b\xi_{\varphi}\rangle= \langle\Psi(b^{*}Ta)\xi_{\varphi},\xi_{\varphi}\rangle= \langle b^{*}Ta\xi_{\varphi},\xi_{\varphi}\rangle= \langle Ta\xi_{\varphi},b\xi_{\varphi}\rangle. ⟨ Ψ ( T ) a ξ φ , b ξ φ ⟩ = ⟨ Ψ ( b ∗ T a ) ξ φ , ξ φ ⟩ = ⟨ b ∗ T a ξ φ , ξ φ ⟩ = ⟨ T a ξ φ , b ξ φ ⟩ .
Since M ξ φ M\xi_{\varphi} M ξ φ is dense in L 2 ( M ) L^{2}(M) L 2 ( M ) , we have Ψ ( T ) = T \Psi(T)=T Ψ ( T ) = T for all T ∈ C T\in\mathcal{C} T ∈ C . Hence, Ψ = id C \Psi=\operatorname{id}_{\mathcal{C}} Ψ = id C . □
Weak Dixmier properties via Poisson boundaries
In this section, we use Poisson boundaries to give a sufficient condition for a W ∗ W^{*} W ∗ -inclusion M ⊂ A M \subset\mathcal{A} M ⊂ A to have the weak Dixmier property. This relates the weak Dixmier property to Poisson boundaries and the associated inner ucp maps. We begin by proving Theorem F.
Theorem 4.1 (Theorem F). Let M ⊂ A M \subset\mathcal{A} M ⊂ A be a W ∗ W^{*} W ∗ -inclusion such that M M M is properly infinite. Then M ⊂ A M \subset\mathcal{A} M ⊂ A has the weak Dixmier property if and only if it has the inner UCP-Dixmier property. Here A \mathcal{A} A need not have separable predual.
Proof. We use a dilation argument similar to that of Haagerup [19 Proposition 2.1 and the proof of Theorem 2.2] , passing from completely positive maps to unitary conjugations by using proper infiniteness. We only need to prove the if direction, since every finite convex combination of unitary conjugations is an M M M -inner ucp map. Assume that M ⊂ A M \subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property.
Because M M M is properly infinite, it contains an infinite sequence of mutually orthogonal projections ( p k ) k = 1 ∞ (p_k)_{k=1}^{\infty} ( p k ) k = 1 ∞ such that p k ∼ 1 M p_k \sim1_M p k ∼ 1 M for all k k k and ∑ k = 1 ∞ p k = 1 M \sum_{k=1}^{\infty}p_k=1_M ∑ k = 1 ∞ p k = 1 M . Take isometries v k ∈ M v_k \in M v k ∈ M with v k v k ∗ = p k v_kv_k^{*}=p_k v k v k ∗ = p k . By this partition of unity, we have a spatial isomorphism
π : A ⟶ ∼ A ⊗ ‾ B ( H ) , π ( a ) = ( v i ∗ a v j ) i , j ≥ 1 , \pi:\mathcal{A}\overset{\sim}{\longrightarrow}\mathcal{A}\overline{\otimes}B(\mathcal{H}), \qquad\pi(a)=(v_i^{*}av_j)_{i,j\geq1}, π : A ⟶ ∼ A ⊗ B ( H ) , π ( a ) = ( v i ∗ a v j ) i , j ≥ 1 ,
where H = ℓ 2 ( N ) \mathcal{H}=\ell^{2}(\mathbb{N}) H = ℓ 2 ( N ) . Its restriction maps M M M onto M ⊗ ‾ B ( H ) M\overline{\otimes}B(\mathcal{H}) M ⊗ B ( H ) and p k p_k p k to 1 M ⊗ e k k 1_M\otimes e_{kk} 1 M ⊗ e k k , where ( e i j ) i , j ≥ 1 (e_{ij})_{i,j\geq1} ( e ij ) i , j ≥ 1 are the standard matrix units of B ( H ) B(\mathcal{H}) B ( H ) . Indeed, on a Hilbert space on which A \mathcal{A} A acts, the unitary ξ ⊗ e k ↦ v k ξ \xi\otimes e_k\mapsto v_k\xi ξ ⊗ e k ↦ v k ξ implements this isomorphism. Hence we identify the inclusion M ⊂ A M\subset\mathcal{A} M ⊂ A with
M ⊗ ‾ B ( H ) ⊂ A ⊗ ‾ B ( H ) . M\overline{\otimes}B(\mathcal{H})\subset\mathcal{A}\overline{\otimes}B(\mathcal{H}). M ⊗ B ( H ) ⊂ A ⊗ B ( H ) .
We only need to prove the weak Dixmier property for an arbitrary element x ∈ A ⊗ ‾ B ( H ) x\in\mathcal{A}\overline{\otimes}B(\mathcal{H}) x ∈ A ⊗ B ( H ) . Write
K x = conv ‾ w { U x U ∗ ∣ U ∈ U ( M ⊗ ‾ B ( H ) ) } . K_x=\overline{\operatorname{conv}}^{\,w}\{UxU^{*}\mid U\in\mathcal{U}(M\overline{\otimes}B(\mathcal{H}))\}. K x = conv w { U x U ∗ ∣ U ∈ U ( M ⊗ B ( H ))} .
Since B ( H ) B(\mathcal{H}) B ( H ) is hyperfinite, take an increasing sequence of finite-dimensional subalgebras F n ⊂ B ( H ) F_n\subset B(\mathcal{H}) F n ⊂ B ( H ) with ⋁ n F n = B ( H ) \bigvee_nF_n=B(\mathcal{H}) ⋁ n F n = B ( H ) . For a free ultrafilter ω \omega ω on N \mathbb{N} N , by normalized Haar averaging we have
a ⊗ 1 B ( H ) : = wo-lim n → ω ∫ U ( F n ) ( 1 M ⊗ u ) x ( 1 M ⊗ u ∗ ) d u ∈ K x ( a ∈ A ) . a\otimes1_{B(\mathcal{H})}:=\operatorname{wo-lim}_{n\to\omega}\int_{\mathcal{U}(F_n)}(1_M\otimes u)x(1_M\otimes u^{*})\,du\in K_x \qquad(a\in\mathcal{A}). a ⊗ 1 B ( H ) := wo-lim n → ω ∫ U ( F n ) ( 1 M ⊗ u ) x ( 1 M ⊗ u ∗ ) d u ∈ K x ( a ∈ A ) .
Since K x K_x K x is weakly closed, convex and invariant under these unitary conjugations, it also contains the weakly closed convex hull of the unitary orbit of a ⊗ 1 B ( H ) a\otimes1_{B(\mathcal{H})} a ⊗ 1 B ( H ) .
We only need to prove that for any M M M -inner ucp map
P ( a ) = ∑ k = 1 ∞ z k a z k ∗ , z k ∈ M , ∑ k = 1 ∞ z k z k ∗ = 1 M , \mathcal{P}(a)=\sum_{k=1}^{\infty}z_ka z_k^{*}, \qquad z_k\in M,\qquad\sum_{k=1}^{\infty}z_kz_k^{*}=1_M, P ( a ) = k = 1 ∑ ∞ z k a z k ∗ , z k ∈ M , k = 1 ∑ ∞ z k z k ∗ = 1 M ,
the element P ( a ) ⊗ 1 B ( H ) \mathcal{P}(a)\otimes1_{B(\mathcal{H})} P ( a ) ⊗ 1 B ( H ) is contained in the weak closure of the convex hull of the unitary orbit of a ⊗ 1 B ( H ) a\otimes1_{B(\mathcal{H})} a ⊗ 1 B ( H ) . Fix an integer n ≥ 1 n \ge1 n ≥ 1 . We construct a partial isometry Z n ∈ M ⊗ ‾ B ( H ) Z_n \in M \overline{\otimes} B(H) Z n ∈ M ⊗ B ( H ) structured as a block matrix where each block is of size n × n n \times n n × n . The first block-row consists of a zero block followed by scaled identity matrices z k I n z_k I_n z k I n , and all subsequent block-rows are identically zero:
Z n = ( 0 n × n z 1 I n z 2 I n z 3 I n ⋯ 0 n × n 0 n × n 0 n × n 0 n × n ⋯ 0 n × n 0 n × n 0 n × n 0 n × n ⋯ ⋮ ⋮ ⋮ ⋮ ⋱ ) . Z_n =
\begin{pmatrix}
0_{n\times n} & z_1I_n & z_2I_n & z_3I_n & \cdots\\
0_{n\times n} & 0_{n\times n} & 0_{n\times n} & 0_{n\times n} & \cdots\\
0_{n\times n} & 0_{n\times n} & 0_{n\times n} & 0_{n\times n} & \cdots\\
\vdots& \vdots& \vdots& \vdots& \ddots
\end{pmatrix}. Z n = 0 n × n 0 n × n 0 n × n ⋮ z 1 I n 0 n × n 0 n × n ⋮ z 2 I n 0 n × n 0 n × n ⋮ z 3 I n 0 n × n 0 n × n ⋮ ⋯ ⋯ ⋯ ⋱ .
The strong convergence of ∑ k = 1 ∞ z k z k ∗ = 1 M \sum_{k=1}^{\infty} z_kz_k^*=1_M ∑ k = 1 ∞ z k z k ∗ = 1 M makes this row a bounded operator. Let E n = ∑ j = 1 n e j j E_n=\sum_{j=1}^n e_{jj} E n = ∑ j = 1 n e j j . Then
Z n Z n ∗ = 1 M ⊗ E n . Z_nZ_n^*=1_M\otimes E_n. Z n Z n ∗ = 1 M ⊗ E n .
Hence, Z n ∗ Z_n^* Z n ∗ is a partial isometry with initial projection 1 M ⊗ E n 1_M\otimes E_n 1 M ⊗ E n . Because the first block-column of Z n Z_n Z n is identically zero, its final projection Z n ∗ Z n Z_n^*Z_n Z n ∗ Z n is orthogonal to 1 M ⊗ E n 1_M\otimes E_n 1 M ⊗ E n . Since M M M is properly infinite, each 1 M ⊗ e j j 1_M\otimes e_{jj} 1 M ⊗ e j j is equivalent to 1 M ⊗ ‾ B ( H ) 1_{M\overline{\otimes}B(H)} 1 M ⊗ B ( H ) . Hence
1 − Z n ∗ Z n ≥ 1 M ⊗ E n ≥ 1 M ⊗ e 11 ∼ 1 M ⊗ ‾ B ( H ) . 1-Z_n^*Z_n \ge1_M\otimes E_n \ge1_M\otimes e_{11}\sim1_{M\overline{\otimes}B(H)}. 1 − Z n ∗ Z n ≥ 1 M ⊗ E n ≥ 1 M ⊗ e 11 ∼ 1 M ⊗ B ( H ) .
We also have
1 − ( 1 M ⊗ E n ) ≥ 1 M ⊗ e n + 1 , n + 1 ∼ 1 M ⊗ ‾ B ( H ) . 1-(1_M\otimes E_n)\ge1_M\otimes e_{n+1,n+1}\sim1_{M\overline{\otimes}B(H)}. 1 − ( 1 M ⊗ E n ) ≥ 1 M ⊗ e n + 1 , n + 1 ∼ 1 M ⊗ B ( H ) .
Therefore, the orthogonal complements of both projections are Murray–von Neumann equivalent to the identity:
1 − ( 1 M ⊗ E n ) ∼ 1 − Z n ∗ Z n ∼ 1 M ⊗ ‾ B ( H ) . 1-(1_M\otimes E_n)\sim1-Z_n^*Z_n\sim1_{M\overline{\otimes}B(H)}. 1 − ( 1 M ⊗ E n ) ∼ 1 − Z n ∗ Z n ∼ 1 M ⊗ B ( H ) .
Hence there exists a partial isometry W n ∈ M ⊗ ‾ B ( H ) W_n\in M\overline{\otimes}B(H) W n ∈ M ⊗ B ( H ) with
W n ∗ W n = 1 − Z n ∗ Z n , W n W n ∗ = 1 − ( 1 M ⊗ E n ) . W_n^*W_n=1-Z_n^*Z_n,\qquad W_nW_n^*=1-(1_M\otimes E_n). W n ∗ W n = 1 − Z n ∗ Z n , W n W n ∗ = 1 − ( 1 M ⊗ E n ) .
Define U n = Z n + W n U_n=Z_n+W_n U n = Z n + W n . Since the initial and final projections of Z n Z_n Z n and W n W_n W n are respectively orthogonal, all mixed terms vanish. Therefore,
U n ∗ U n = Z n ∗ Z n + W n ∗ W n = 1 , U n U n ∗ = Z n Z n ∗ + W n W n ∗ = 1. U_n^*U_n=Z_n^*Z_n+W_n^*W_n=1,\qquad U_nU_n^*=Z_nZ_n^*+W_nW_n^*=1. U n ∗ U n = Z n ∗ Z n + W n ∗ W n = 1 , U n U n ∗ = Z n Z n ∗ + W n W n ∗ = 1.
Hence we complete the partial isometry into a unitary U n ∈ U ( M ⊗ ‾ B ( H ) ) U_n\in U(M\overline{\otimes}B(H)) U n ∈ U ( M ⊗ B ( H )) . Since ( 1 M ⊗ E n ) W n = 0 (1_M\otimes E_n)W_n=0 ( 1 M ⊗ E n ) W n = 0 , the first block-row of U n U_n U n remains the first block-row of Z n Z_n Z n . In block-row notation, we write
U n = ( Z n W n ) . U_n=\begin{pmatrix} Z_n\\ W_n\end{pmatrix}. U n = ( Z n W n ) .
We conjugate a ⊗ 1 B ( H ) a\otimes1_{B(H)} a ⊗ 1 B ( H ) by U n U_n U n . Since
Z n ( a ⊗ 1 B ( H ) ) Z n ∗ = P ( a ) ⊗ E n , Z_n(a\otimes1_{B(H)})Z_n^*=P(a)\otimes E_n, Z n ( a ⊗ 1 B ( H ) ) Z n ∗ = P ( a ) ⊗ E n ,
we have, with respect to the first n n n coordinates and their orthogonal complement,
U n ( a ⊗ 1 B ( H ) ) U n ∗ = ( P ( a ) I n ∗ ∗ ∗ ) . (22) U_n(a\otimes1_{B(H)})U_n^*=
\begin{pmatrix}
P(a)I_n & *\\
* & *
\end{pmatrix}. \tag*{(22)} U n ( a ⊗ 1 B ( H ) ) U n ∗ = ( P ( a ) I n ∗ ∗ ∗ ) . (22)
Let b n = U n ( a ⊗ 1 B ( H ) ) U n ∗ − P ( a ) ⊗ E n b_n=U_n(a\otimes1_{B(H)})U_n^*-P(a)\otimes E_n b n = U n ( a ⊗ 1 B ( H ) ) U n ∗ − P ( a ) ⊗ E n . By (22) , we have ( 1 A ⊗ E n ) b n ( 1 A ⊗ E n ) = 0 (1_A\otimes E_n)b_n(1_A\otimes E_n)=0 ( 1 A ⊗ E n ) b n ( 1 A ⊗ E n ) = 0 . Since ∥ b n ∥ ≤ 2 ∥ a ∥ \lVert b_n\rVert\le2\lVert a\rVert ∥ b n ∥ ≤ 2 ∥ a ∥ and E n ↗ 1 B ( H ) E_n\nearrow1_{B(H)} E n ↗ 1 B ( H ) strongly, we have
b n = ( 1 − 1 A ⊗ E n ) b n + ( 1 A ⊗ E n ) b n ( 1 − 1 A ⊗ E n ) ⟶ 0 b_n=(1-1_A\otimes E_n)b_n+(1_A\otimes E_n)b_n(1-1_A\otimes E_n)\longrightarrow0 b n = ( 1 − 1 A ⊗ E n ) b n + ( 1 A ⊗ E n ) b n ( 1 − 1 A ⊗ E n ) ⟶ 0
weakly ( n → ∞ ) (n\to\infty) ( n → ∞ ) . Therefore,
w o - l i m n → ∞ U n ( a ⊗ 1 B ( H ) ) U n ∗ = P ( a ) ⊗ 1 B ( H ) . \mathop{\mathrm{wo\text{-}lim}}_{n\to\infty}U_n(a\otimes1_{B(H)})U_n^*=P(a)\otimes1_{B(H)}. wo - lim n → ∞ U n ( a ⊗ 1 B ( H ) ) U n ∗ = P ( a ) ⊗ 1 B ( H ) .
Because this holds for any M M M -inner ucp map P P P , the entire set
{ P ( a ) ⊗ 1 B ( H ) ∣ P is an M -inner ucp map } \{P(a)\otimes1_{B(H)}\mid P\text{ is an }M\text{-inner ucp map}\} { P ( a ) ⊗ 1 B ( H ) ∣ P is an M -inner ucp map }
is contained in K x K_x K x .
By the assumption that the original inclusion M ⊂ A M \subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property, there exists an element
y ∈ { P ( a ) ∣ P is an M -inner ucp map } ‾ w ∩ ( M ′ ∩ A ) . y \in\overline{\{P(a)\mid P\text{ is an }M\text{-inner ucp map}\}}^{w} \cap(M' \cap\mathcal{A}). y ∈ { P ( a ) ∣ P is an M -inner ucp map } w ∩ ( M ′ ∩ A ) .
Because the mapping a ↦ a ⊗ 1 B ( H ) a \mapsto a\otimes1_{B(\mathcal{H})} a ↦ a ⊗ 1 B ( H ) is normal, the element y ⊗ 1 B ( H ) y\otimes1_{B(\mathcal{H})} y ⊗ 1 B ( H ) lies in the weak closure of { P ( a ) ⊗ 1 B ( H ) } \{P(a)\otimes1_{B(\mathcal{H})}\} { P ( a ) ⊗ 1 B ( H ) } . Since K x K_x K x is already weakly closed, it contains y ⊗ 1 B ( H ) y\otimes1_{B(\mathcal{H})} y ⊗ 1 B ( H ) . Hence, y ⊗ 1 B ( H ) y\otimes1_{B(\mathcal{H})} y ⊗ 1 B ( H ) belongs to
( M ⊗ ‾ B ( H ) ) ′ ∩ ( A ⊗ ‾ B ( H ) ) = ( M ′ ∩ A ) ⊗ ‾ C 1 B ( H ) (M\overline{\otimes}B(\mathcal{H}))' \cap(\mathcal{A}\overline{\otimes}B(\mathcal{H})) = (M' \cap\mathcal{A})\overline{\otimes}\mathbb{C}1_{B(\mathcal{H})} ( M ⊗ B ( H ) ) ′ ∩ ( A ⊗ B ( H )) = ( M ′ ∩ A ) ⊗ C 1 B ( H )
and lies in the weakly closed convex hull of the unitary orbit of x x x , which finishes the proof. □
Combining Theorem 4.1 with Corollary 3.14, we have the weak Dixmier property for inclusions that map into a Poisson boundary.
Theorem 4.2 (Theorem D ). Let M M M be a von Neumann algebra with separable predual and φ \varphi φ be an nrsg hyperstate on B ( L 2 ( M ) ) B(L^{2}(M)) B ( L 2 ( M )) with the Poisson boundary B φ B_{\varphi} B φ . Assume that one of the following conditions holds:
(i) M M M is properly infinite, i.e., M = M I ∞ ⊕ M I I ∞ ⊕ M I I I M=M_{\mathrm{I}_{\infty}}\oplus M_{\mathrm{II}_{\infty}}\oplus M_{\mathrm{III}} M = M I ∞ ⊕ M II ∞ ⊕ M III ;
(ii) M M M is finite with normal faithful trace τ \tau τ and φ = ∫ U ( M ) ⟨ ⋅ u ξ τ , u ξ τ ⟩ d μ ( u ) \varphi=\int_{\mathcal{U}(M)}\langle\cdot u\xi_{\tau},u\xi_{\tau}\rangle\,\mathrm{d}\mu(u) φ = ∫ U ( M ) ⟨ ⋅ u ξ τ , u ξ τ ⟩ d μ ( u ) for some μ ∈ Prob ( U ( M ) ) \mu\in\operatorname{Prob}(\mathcal{U}(M)) μ ∈ Prob ( U ( M )) .
Then any W ∗ W^{*} W ∗ -inclusion M ⊂ A M\subset\mathcal{A} M ⊂ A equipped with a normal faithful ucp map Φ : A → B φ \Phi:\mathcal{A}\to B_{\varphi} Φ : A → B φ with Φ ∣ M = id M \Phi|_{M}=\operatorname{id}_{M} Φ ∣ M = id M has the weak Dixmier property.
Proof. Let M ⊂ A M\subset\mathcal{A} M ⊂ A be a W ∗ W^{*} W ∗ -inclusion equipped with a normal faithful ucp map Φ : A → B φ \Phi:\mathcal{A}\to B_{\varphi} Φ : A → B φ satisfying Φ ∣ M = id M \Phi|_{M}=\operatorname{id}_{M} Φ ∣ M = id M . By Corollary 3.14, M ⊂ A M\subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property.
We consider the two cases separately.
Case (i): Assume M M M is properly infinite. By Theorem 4.1, for properly infinite von Neumann algebras, the weak Dixmier property is equivalent to the inner UCP-Dixmier property. Since M ⊂ A M\subset\mathcal{A} M ⊂ A already has the inner UCP-Dixmier property by Corollary 3.14, it has the weak Dixmier property.
Case (ii): Assume M M M is finite with trace τ \tau τ , and φ = ∫ U ( M ) ⟨ ⋅ u ξ τ , u ξ τ ⟩ d μ ( u ) \varphi=\int_{\mathcal{U}(M)}\langle\cdot u\xi_{\tau},u\xi_{\tau}\rangle\,\mathrm{d}\mu(u) φ = ∫ U ( M ) ⟨ ⋅ u ξ τ , u ξ τ ⟩ d μ ( u ) for some probability measure μ ∈ Prob ( U ( M ) ) \mu\in\operatorname{Prob}(\mathcal{U}(M)) μ ∈ Prob ( U ( M )) . In this setting, the associated ucp map Q : A → A Q:\mathcal{A}\to\mathcal{A} Q : A → A is given by Q ( x ) = ∫ U ( M ) u ∗ x u d μ ( u ) Q(x)=\int_{\mathcal{U}(M)}u^{*}xu\,\mathrm{d}\mu(u) Q ( x ) = ∫ U ( M ) u ∗ xu d μ ( u ) for all x ∈ A x\in\mathcal{A} x ∈ A . The adjoint direction follows from the vector-state convention φ = ∫ ω u ξ τ d μ ( u ) \varphi=\int\omega_{u\xi_{\tau}}\,\mathrm{d}\mu(u) φ = ∫ ω u ξ τ d μ ( u ) . Because Q Q Q is an integral of unitary conjugations, the elements Q k ( x ) Q^{k}(x) Q k ( x ) and their Cesàro means C n ( x ) = ( 1 / n ) ∑ k = 0 n − 1 Q k ( x ) C_{n}(x)=(1/n)\sum_{k=0}^{n-1}Q^{k}(x) C n ( x ) = ( 1/ n ) ∑ k = 0 n − 1 Q k ( x ) lie in the weakly closed convex hull of the unitary orbit of x x x :
C n ( x ) ∈ conv ‾ w { u x u ∗ ∣ u ∈ U ( M ) } . C_{n}(x)\in\overline{\operatorname{conv}}^{w}\{uxu^{*}\mid u\in\mathcal{U}(M)\}. C n ( x ) ∈ conv w { ux u ∗ ∣ u ∈ U ( M )} .
By the proof of Corollary 3.14, any weak limit point z z z of the sequence ( C n ( x ) ) n = 1 ∞ (C_{n}(x))_{n=1}^{\infty} ( C n ( x ) ) n = 1 ∞ is an element of M ′ ∩ A M'\cap\mathcal{A} M ′ ∩ A . Since the weakly closed convex hull of the unitary orbit is weakly closed, the limit point z z z remains within it. Hence
z ∈ conv ‾ w { u x u ∗ ∣ u ∈ U ( M ) } ∩ ( M ′ ∩ A ) ≠ ∅ . z\in\overline{\operatorname{conv}}^{w}\{uxu^{*}\mid u\in\mathcal{U}(M)\}\cap(M'\cap\mathcal{A})\ne\varnothing. z ∈ conv w { ux u ∗ ∣ u ∈ U ( M )} ∩ ( M ′ ∩ A ) = ∅ .
This proves that M ⊂ A M\subset\mathcal{A} M ⊂ A has the weak Dixmier property. □
As a first application, we recover the weak Dixmier property for inclusions with expectation.
Corollary 4.3 (Corollary E , [39 , 31 ] ). Let M ⊂ A M\subset\mathcal{A} M ⊂ A be an inclusion of countably decomposable von Neumann algebras with a normal faithful conditional expectation. Then M ⊂ A M\subset\mathcal{A} M ⊂ A has the weak Dixmier property.
Proof. Let E : A → M E:\mathcal{A}\to M E : A → M be the given normal faithful conditional expectation. As in [31 p. 122, proof of the main theorem] , take an increasing net ( M i ) i ∈ I (M_{i})_{i\in I} ( M i ) i ∈ I of von Neumann subalgebras with separable preduals and normal faithful conditional expectations E M i : M → M i E_{M_{i}}:M\to M_{i} E M i : M → M i , such that ⋁ i M i = M \bigvee_{i}M_{i}=M ⋁ i M i = M . For x ∈ A x\in\mathcal{A} x ∈ A , let
K x i = conv ‾ w { u x u ∗ ∣ u ∈ U ( M i ) } . (23) K_{x}^{i}=\overline{\operatorname{conv}}^{w}\{uxu^{*}\mid u\in\mathcal{U}(M_{i})\}. \tag*{(23)} K x i = conv w { ux u ∗ ∣ u ∈ U ( M i )} . (23)
We only need to prove that K x i ∩ M i ′ ≠ ∅ K_x^i \cap M_i' \ne\varnothing K x i ∩ M i ′ = ∅ for every i i i : by weak compactness, any choice y i ∈ K x i ∩ M i ′ y_i \in K_x^i \cap M_i' y i ∈ K x i ∩ M i ′ has a weak cluster point in
conv ‾ w { u x u ∗ ∣ u ∈ U ( M ) } ∩ ( M ′ ∩ A ) . \overline{\operatorname{conv}}^{\,w}\{uxu^* \mid u \in\mathcal{U}(M)\} \cap(M' \cap A). conv w { ux u ∗ ∣ u ∈ U ( M )} ∩ ( M ′ ∩ A ) .
Since E M i ∘ E : A → M i E_{M_i} \circ E : A \to M_i E M i ∘ E : A → M i is a normal faithful conditional expectation, we may therefore assume that M M M has separable predual. No separability assumption on A A A is needed in the argument below or in the auxiliary averaging results used here. Decompose M = M f i n ⊕ M ∞ M = M_{\mathrm{fin}} \oplus M_{\infty} M = M fin ⊕ M ∞ into its finite and properly infinite parts, omitting zero summands, and write p 1 , p 2 p_1,p_2 p 1 , p 2 for their units. Averaging by p 1 + e 2 π i t p 2 p_1 + e^{2\pi it}p_2 p 1 + e 2 π i t p 2 removes the off-diagonal terms of every x ∈ A x \in A x ∈ A :
∫ 0 1 ( p 1 + e 2 π i t p 2 ) x ( p 1 + e 2 π i t p 2 ) ∗ d t = p 1 x p 1 + p 2 x p 2 . (24) \int_0^1 (p_1+e^{2\pi it}p_2)x(p_1+e^{2\pi it}p_2)^*\,dt = p_1xp_1+p_2xp_2. \tag*{(24)} ∫ 0 1 ( p 1 + e 2 π i t p 2 ) x ( p 1 + e 2 π i t p 2 ) ∗ d t = p 1 x p 1 + p 2 x p 2 . (24)
The restrictions E i : p i A p i → p i M E_i : p_iAp_i \to p_iM E i : p i A p i → p i M are normal faithful conditional expectations for i = 1 , 2 i=1,2 i = 1 , 2 .
Take a normal faithful trace τ \tau τ on M f i n M_{\mathrm{fin}} M fin , a countable adjoint-closed family of unitaries ( u k ) (u_k) ( u k ) generating M f i n M_{\mathrm{fin}} M fin , and c k > 0 c_k > 0 c k > 0 with ∑ k c k = 1 \sum_k c_k = 1 ∑ k c k = 1 . The state
φ 1 = ∑ k c k ⟨ ⋅ u k ξ τ , u k ξ τ ⟩ \varphi_1 = \sum_k c_k \langle\,\cdot\, u_k\xi_\tau,u_k\xi_\tau\rangle φ 1 = k ∑ c k ⟨ ⋅ u k ξ τ , u k ξ τ ⟩
is normal and regular. Both strong generation conditions follow from the weak density of C [ ( u k ) ] \mathbb{C}[(u_k)] C [( u k )] , since ( u k ) (u_k) ( u k ) is adjoint-closed. Composing E 1 E_1 E 1 with M f i n ⊂ B φ 1 M_{\mathrm{fin}} \subset B_{\varphi_1} M fin ⊂ B φ 1 , by Theorem 4.2 (ii), we have that M f i n ⊂ p 1 A p 1 M_{\mathrm{fin}} \subset p_1Ap_1 M fin ⊂ p 1 A p 1 has the weak Dixmier property.
For M ∞ M_\infty M ∞ , take an nrsg hyperstate φ 2 \varphi_2 φ 2 by Example 3.9 . Composing E 2 E_2 E 2 with M ∞ ⊂ B φ 2 M_\infty\subset B_{\varphi_2} M ∞ ⊂ B φ 2 and applying Theorem 4.2 (i), we have the same conclusion for this corner. Taking product nets of the two unitary averaging nets after applying (24) , we have an element of M ′ ∩ A M' \cap A M ′ ∩ A in the closed convex unitary orbit of x x x . □
As a semifinite analogue of Corollary 4.3, Isono [Iso25, Theorem A] proved that, for a W*-inclusion M ⊂ A M \subset A M ⊂ A with a normal faithful semifinite operator-valued weight T : A + → M ^ + T : A_+ \to\widehat{M}_+ T : A + → M + , every x ∈ A + x \in A_+ x ∈ A + with T ( x ) T(x) T ( x ) bounded satisfies
conv ‾ w { u x u ∗ ∣ u ∈ U ( M ) } ∩ ( M ′ ∩ A ) ≠ ∅ . \overline{\operatorname{conv}}^{\,w}\{uxu^* \mid u \in\mathcal{U}(M)\} \cap(M' \cap A) \ne\varnothing. conv w { ux u ∗ ∣ u ∈ U ( M )} ∩ ( M ′ ∩ A ) = ∅ .
To generalize Isono’s result, we first introduce a generalization of such weights.
Definition 4.4. Let M ⊂ A M \subset A M ⊂ A and M ⊂ B M \subset B M ⊂ B be W*-inclusions. A map T : A + → B ^ + T : A_+ \to\widehat{B}_+ T : A + → B + is called an M M M -bimodular semifinite positive map if it is additive, positively homogeneous and satisfies
T ( a ∗ x a ) = a ∗ T ( x ) a ( x ∈ A + , a ∈ M ) . T(a^*xa)=a^*T(x)a \qquad(x \in A_+,\, a \in M). T ( a ∗ x a ) = a ∗ T ( x ) a ( x ∈ A + , a ∈ M ) . We require normality, meaning preservation of increasing bounded nets, faithfulness, and ultraweak density of
m T + = { x ∈ A + ∣ T ( x ) ∈ B + } . \mathfrak{m}_T^+ = \{x \in A_+ \mid T(x) \in B_+\}. m T + = { x ∈ A + ∣ T ( x ) ∈ B + } . Here B ^ + \widehat{B}_+ B + is the extended positive cone, and T ( x ) < ∞ T(x)<\infty T ( x ) < ∞ will always mean that T ( x ) T(x) T ( x ) is bounded. For every ρ ∈ B ∗ + \rho\in B_*^+ ρ ∈ B ∗ + , the function ρ ∘ T \rho\circ T ρ ∘ T is a normal weight. In particular, its sublevel sets are ultraweakly closed. It follows that, for C < ∞ C<\infty C < ∞ , the set
{ x ∈ A + ∣ T ( x ) ≤ C 1 } (25) \{x \in A_+ \mid T(x) \le C1\} \tag*{(25)} { x ∈ A + ∣ T ( x ) ≤ C 1 } (25) is ultraweakly closed. We use this scalar interpretation of lower semicontinuity throughout, see [18 Sections 1–2] .
The following lemma is a semifinite analogue of Lemma 3.11 . When the invariant state is replaced by an invariant semifinite weight, Cesàro limits still satisfy the following multiplicativity.
Lemma 4.5. Let ψ A \psi_A ψ A be a normal faithful semifinite weight on a von Neumann algebra A A A , and let P : A → A \mathcal{P}: A \to A P : A → A be a normal ucp map satisfying ψ A ∘ P = ψ A \psi_A \circ\mathcal{P} = \psi_A ψ A ∘ P = ψ A . Let C n = ( 1 / n ) ∑ k = 0 n − 1 P k C_n = (1/n)\sum_{k=0}^{n-1}\mathcal{P}^k C n = ( 1/ n ) ∑ k = 0 n − 1 P k . Fix a free ultrafilter ω \omega ω and define
E ( x ) = wo-lim n → ω C n ( x ) . E(x) = \operatorname{wo\text{-}lim}_{n\to\omega} C_n(x). E ( x ) = wo - lim n → ω C n ( x ) . If x ∈ A + x \in A_+ x ∈ A + and ψ A ( x ) < ∞ \psi_A(x) < \infty ψ A ( x ) < ∞ , then P ( E ( x ) ) = E ( x ) \mathcal{P}(E(x)) = E(x) P ( E ( x )) = E ( x ) and P ( E ( x ) 2 ) = E ( x ) 2 \mathcal{P}(E(x)^2) = E(x)^2 P ( E ( x ) 2 ) = E ( x ) 2 .
Proof. Let y = E ( x ) y = E(x) y = E ( x ) . Since ∥ P ( C n ( x ) ) − C n ( x ) ∥ ≤ 2 ∥ x ∥ / n \|\mathcal{P}(C_n(x)) - C_n(x)\| \leq2\|x\|/n ∥ P ( C n ( x )) − C n ( x ) ∥ ≤ 2∥ x ∥/ n , by normality of P \mathcal{P} P , we have P ( y ) = y \mathcal{P}(y) = y P ( y ) = y . By lower semicontinuity and invariance of the weight, we have
ψ A ( y ) ≤ ψ A ( x ) < ∞ , ψ A ( y 2 ) ≤ ∥ y ∥ ψ A ( y ) < ∞ . \psi_A(y) \leq\psi_A(x) < \infty,\qquad\psi_A(y^2) \leq\|y\|\psi_A(y) < \infty. ψ A ( y ) ≤ ψ A ( x ) < ∞ , ψ A ( y 2 ) ≤ ∥ y ∥ ψ A ( y ) < ∞.
By Kadison’s inequality, P ( y 2 ) − y 2 ≥ 0 \mathcal{P}(y^2) - y^2 \geq0 P ( y 2 ) − y 2 ≥ 0 . Since ψ A ∘ P = ψ A \psi_A \circ\mathcal{P} = \psi_A ψ A ∘ P = ψ A and ψ A ( y 2 ) < ∞ \psi_A(y^2) < \infty ψ A ( y 2 ) < ∞ ,
ψ A ( P ( y 2 ) − y 2 ) = ψ A ( P ( y 2 ) ) − ψ A ( y 2 ) = 0. \psi_A(\mathcal{P}(y^2) - y^2) = \psi_A(\mathcal{P}(y^2)) - \psi_A(y^2) = 0. ψ A ( P ( y 2 ) − y 2 ) = ψ A ( P ( y 2 )) − ψ A ( y 2 ) = 0.
By faithfulness, we have P ( y 2 ) = y 2 \mathcal{P}(y^2) = y^2 P ( y 2 ) = y 2 .
By this lemma, we have the following semifinite version of Proposition 3.13.
Theorem 4.6. Let M ⊂ A M \subset A M ⊂ A be a W ∗ W^* W ∗ -inclusion. Assume that M M M has separable predual, and let φ \varphi φ be an nrsg M M M -hyperstate. Assume that there exists a normal faithful semifinite M M M -bimodular positive map T : A + → B ^ φ + T: A_+ \to\widehat{B}_{\varphi+} T : A + → B φ + in the preceding sense. Define the normal M M M -inner ucp map Q : A → A Q: A \to A Q : A → A by Q ( x ) = ∑ k y k x y k ∗ Q(x) = \sum_k y_kxy_k^* Q ( x ) = ∑ k y k x y k ∗ . For any x ∈ A + x \in A_+ x ∈ A + with T ( x ) < ∞ T(x) < \infty T ( x ) < ∞ , every ultraweak cluster point of the Cesàro means C n ( x ) = ( 1 / n ) ∑ k = 0 n − 1 Q k ( x ) C_n(x) = (1/n)\sum_{k=0}^{n-1}Q^k(x) C n ( x ) = ( 1/ n ) ∑ k = 0 n − 1 Q k ( x ) belongs to M ′ ∩ A M' \cap A M ′ ∩ A .
Proof. Let ζ \zeta ζ be the canonical normal faithful state on B φ B_\varphi B φ and let ψ A = ζ ∘ T \psi_A = \zeta\circ T ψ A = ζ ∘ T . This is a normal faithful weight. It is semifinite because its finite positive domain contains the ultraweakly dense cone m T + m_T^+ m T + .
By normality and M M M -bimodularity, in the extended positive cone we have
T ( Q ( x ) ) = ∑ k y k T ( x ) y k ∗ = Q φ ( T ( x ) ) . T(Q(x)) = \sum_k y_kT(x)y_k^* = Q_\varphi(T(x)). T ( Q ( x )) = k ∑ y k T ( x ) y k ∗ = Q φ ( T ( x )) .
By (20) , ζ ∘ Q φ = ζ \zeta\circ Q_\varphi= \zeta ζ ∘ Q φ = ζ on ( B φ ) + (B_\varphi)_+ ( B φ ) + . This identity extends to the extended positive cone by monotone approximation. Hence ψ A ∘ Q = ψ A \psi_A \circ Q = \psi_A ψ A ∘ Q = ψ A .
Let a a a be an ultraweak cluster point of ( C n ( x ) ) (C_n(x)) ( C n ( x )) . By the proof of Lemma 4.5 , applied along a convergent subnet, we have Q ( a ) = a Q(a) = a Q ( a ) = a and Q ( a 2 ) = a 2 Q(a^2) = a^2 Q ( a 2 ) = a 2 . Hence,
∑ k ( y k a − a y k ) ( y k a − a y k ) ∗ = Q ( a 2 ) − Q ( a ) a − a Q ( a ) + a 2 = 0. \sum_k (y_ka - ay_k)(y_ka - ay_k)^* = Q(a^2) - Q(a)a - aQ(a) + a^2 = 0. k ∑ ( y k a − a y k ) ( y k a − a y k ) ∗ = Q ( a 2 ) − Q ( a ) a − a Q ( a ) + a 2 = 0.
Hence [ a , y k ] = 0 [a,y_k] = 0 [ a , y k ] = 0 for every k k k . Since C [ ( y k ) ] ‾ w = M \overline{\mathbb{C}[(y_k)]}^{\,w} = M C [( y k )] w = M , a ∈ M ′ ∩ A a \in M' \cap A a ∈ M ′ ∩ A . □
As in Theorem 4.2 , we can convert these Cesàro limits into unitary averages.
Theorem 4.7. Let M M M have separable predual and let φ \varphi φ be an nrsg hyperstate. Assume that one of the following conditions holds:
φ ( a ) = ∫ U ( M ) ⟨ a u ξ τ , u ξ τ ⟩ d μ ( u ) (26) \varphi(a) = \int_{U(M)} \langle au\xi_\tau, u\xi_\tau\rangle\,d\mu(u) \tag*{(26)} φ ( a ) = ∫ U ( M ) ⟨ a u ξ τ , u ξ τ ⟩ d μ ( u ) (26) for some μ ∈ Prob ( U ( M ) ) \mu\in\operatorname{Prob}(U(M)) μ ∈ Prob ( U ( M )) .
Let M ⊂ A M \subset A M ⊂ A be a W ∗ W^* W ∗ -inclusion and let T : A + → B ^ φ + T: A_+ \to\widehat{B}_{\varphi+} T : A + → B φ + be a normal faithful semifinite M M M -bimodular positive map. Then, for x ∈ A + x \in A_+ x ∈ A + with T ( x ) < ∞ T(x) < \infty T ( x ) < ∞ ,
conv ‾ w { u x u ∗ ∣ u ∈ U ( M ) } ∩ ( M ′ ∩ A ) ≠ ∅ . \overline{\operatorname{conv}}^{\,w}\{uxu^* \mid u \in U(M)\} \cap(M' \cap A) \neq\varnothing. conv w { ux u ∗ ∣ u ∈ U ( M )} ∩ ( M ′ ∩ A ) = ∅ .
The relative bicentralizer conjecture
Proof. Let Q Q Q and its Cesàro means C n C_n C n be as in Theorem 4.6 . Let C = ∥ T ( x ) ∥ C=\lVert T(x)\rVert C = ∥ T ( x )∥ and K x = conv ‾ w { u x u ∗ ∣ u ∈ U ( M ) } K_x=\overline{\operatorname{conv}}^{w}\{uxu^*\mid u\in\mathcal{U}(M)\} K x = conv w { ux u ∗ ∣ u ∈ U ( M )} . For u ∈ U ( M ) u\in\mathcal{U}(M) u ∈ U ( M ) , T ( u x u ∗ ) = u T ( x ) u ∗ ≤ C 1 T(uxu^*)=uT(x)u^*\leq C1 T ( ux u ∗ ) = u T ( x ) u ∗ ≤ C 1 . By the ultraweak closedness of (25) ,
T ( b ) ≤ C 1 ( b ∈ K x ) . T(b)\leq C1\qquad(b\in K_x). T ( b ) ≤ C 1 ( b ∈ K x ) .
Suppose first that M M M is properly infinite. Take isometries ( v i ) ⊂ M (v_i)\subset M ( v i ) ⊂ M with v i ∗ v j = δ i j 1 v_i^*v_j=\delta_{ij}1 v i ∗ v j = δ ij 1 and ∑ i v i v i ∗ = 1 \sum_i v_iv_i^*=1 ∑ i v i v i ∗ = 1 , and use the isomorphism
π : A → ∼ A ⊗ ‾ B ( ℓ 2 ( N ) ) , π ( b ) = ∑ i , j v i ∗ b v j ⊗ e i j . \pi:\mathcal{A}\xrightarrow{\sim}\mathcal{A}\mathbin{\overline{\otimes}}B(\ell^2(\mathbb{N})),\qquad\pi(b)=\sum_{i,j}v_i^*bv_j\otimes e_{ij}. π : A ∼ A ⊗ B ( ℓ 2 ( N )) , π ( b ) = i , j ∑ v i ∗ b v j ⊗ e ij .
It maps M M M onto M ⊗ ‾ B ( ℓ 2 ( N ) ) M\mathbin{\overline{\otimes}}B(\ell^2(\mathbb{N})) M ⊗ B ( ℓ 2 ( N )) . As in the proof of Theorem 4.1, by unitary averaging over the type I tensor factor, we have a ⊗ 1 ∈ π ( K x ) a\otimes1\in\pi(K_x) a ⊗ 1 ∈ π ( K x ) for some a ∈ A + a\in\mathcal{A}_+ a ∈ A + . Let b = π − 1 ( a ⊗ 1 ) b=\pi^{-1}(a\otimes1) b = π − 1 ( a ⊗ 1 ) . Since b ∈ K x b\in K_x b ∈ K x and a = v 1 ∗ b v 1 a=v_1^*bv_1 a = v 1 ∗ b v 1 ,
T ( a ) = v 1 ∗ T ( b ) v 1 ≤ C 1. T(a)=v_1^*T(b)v_1\leq C1. T ( a ) = v 1 ∗ T ( b ) v 1 ≤ C 1.
By Theorem 4.6 , an ultraweak cluster point c c c of ( C n ( a ) ) (C_n(a)) ( C n ( a )) belongs to M ′ ∩ A M'\cap\mathcal{A} M ′ ∩ A .
The matrix-dilation argument in the proof of Theorem 4.1 applies to each M M M -inner ucp map. Hence C n ( a ) ⊗ 1 C_n(a)\otimes1 C n ( a ) ⊗ 1 belongs to the weakly closed unitary convex hull of a ⊗ 1 a\otimes1 a ⊗ 1 for every n n n . That hull is contained in π ( K x ) \pi(K_x) π ( K x ) , so c ⊗ 1 ∈ π ( K x ) c\otimes1\in\pi(K_x) c ⊗ 1 ∈ π ( K x ) . Since c ⊗ 1 ∈ ( M ⊗ ‾ B ( ℓ 2 ( N ) ) ) ′ c\otimes1\in(M\mathbin{\overline{\otimes}}B(\ell^2(\mathbb{N})))' c ⊗ 1 ∈ ( M ⊗ B ( ℓ 2 ( N )) ) ′ ,
π − 1 ( c ⊗ 1 ) ∈ K x ∩ ( M ′ ∩ A ) . \pi^{-1}(c\otimes1)\in K_x\cap(M'\cap\mathcal{A}). π − 1 ( c ⊗ 1 ) ∈ K x ∩ ( M ′ ∩ A ) .
Suppose now that M M M is finite and φ \varphi φ satisfies (26) . With the present vector-state convention,
Q ( b ) = ∫ U ( M ) u ∗ b u d μ ( u ) . Q(b)=\int_{\mathcal{U}(M)}u^*bu\,d\mu(u). Q ( b ) = ∫ U ( M ) u ∗ b u d μ ( u ) .
Its iterates and their Cesàro means C n C_n C n preserve K x K_x K x . The ultraweak cluster points of ( C n ( x ) ) (C_n(x)) ( C n ( x )) belong to M ′ ∩ A M'\cap\mathcal{A} M ′ ∩ A by Theorem 4.6 , which proves the assertion. □
Composing an operator-valued weight onto M M M with the boundary inclusion recovers Isono’s result.
Corollary 4.8 ([Iso25]). Let M ⊂ A M\subset\mathcal{A} M ⊂ A be an inclusion of countably decomposable von Neumann algebras, and let T : A + → M ^ + T:\mathcal{A}_+\to\widehat{M}_+ T : A + → M + be a normal faithful semifinite operator-valued weight. For every x ∈ A + x\in\mathcal{A}_+ x ∈ A + such that T ( x ) T(x) T ( x ) is bounded,
conv ‾ w { u x u ∗ ∣ u ∈ U ( M ) } ∩ ( M ′ ∩ A ) ≠ ∅ . \overline{\operatorname{conv}}^{w}\{uxu^*\mid u\in\mathcal{U}(M)\}\cap(M'\cap\mathcal{A})\neq\varnothing. conv w { ux u ∗ ∣ u ∈ U ( M )} ∩ ( M ′ ∩ A ) = ∅ .
Proof. As in the beginning of the proof of Corollary 4.3, we may assume that M M M has separable predual. Write M = M f i n ⊕ M ∞ M=M_{\mathrm{fin}}\oplus M_\infty M = M fin ⊕ M ∞ for its finite and properly infinite parts, with units p 1 , p 2 p_1,p_2 p 1 , p 2 . Averaging over the unitaries p 1 + e 2 π i t p 2 p_1+e^{2\pi it}p_2 p 1 + e 2 π i t p 2 sends x x x to p 1 x p 1 + p 2 x p 2 p_1xp_1+p_2xp_2 p 1 x p 1 + p 2 x p 2 within its unitary convex hull.
The restrictions T i : ( p i A p i ) + → p i M ^ + T_i:(p_i\mathcal{A}p_i)_+\to p_i\widehat{M}_+ T i : ( p i A p i ) + → p i M + are normal, faithful and semifinite. Indeed, compression of the ultraweakly dense finite positive domain of T T T is ultraweakly dense in ( p i A p i ) + (p_i\mathcal{A}p_i)_+ ( p i A p i ) + , and
T i ( p i a p i ) = p i T ( a ) p i . T_i(p_iap_i)=p_iT(a)p_i. T i ( p i a p i ) = p i T ( a ) p i .
Moreover, T i ( p i x p i ) T_i(p_ixp_i) T i ( p i x p i ) is bounded.
On M f i n M_{\mathrm{fin}} M fin take an atomic measure on a countable symmetric family of unitaries generating M f i n M_{\mathrm{fin}} M fin , with every atom having positive mass. Then we have an nrsg hyperstate φ 1 \varphi_1 φ 1 with respect to a normal faithful tracial state. On M ∞ M_\infty M ∞ take an nrsg hyperstate φ 2 \varphi_2 φ 2 by Example 3.9 . Compose each T i T_i T i with the extension to extended positive cones of the canonical embedding p i M ⊂ B φ i p_iM\subset B_{\varphi_i} p i M ⊂ B φ i . These composites satisfy the hypotheses of Theorem 4.7 . No normalization on p i M p_iM p i M is required. Apply that theorem separately to the two positive diagonal blocks.
Finally, block-diagonal unitaries in U ( M f i n ) ⊕ U ( M ∞ ) \mathcal{U}(M_{\mathrm{fin}})\oplus\mathcal{U}(M_\infty) U ( M fin ) ⊕ U ( M ∞ ) belong to U ( M ) \mathcal{U}(M) U ( M ) . Taking product nets of the two averaging nets, we have a limit in M ′ ∩ A M'\cap\mathcal{A} M ′ ∩ A , still in the weakly closed convex hull of the unitary orbit of x x x . □
The relative bicentralizer conjecture
In this section, we prove Theorem A. Fix an inclusion M ⊂ P M \subset P M ⊂ P of countably decomposable von Neumann algebras with a normal faithful conditional expectation E : P → M E:P \to M E : P → M . We only need to prove that M ⊂ c ( P ) M \subset c(P) M ⊂ c ( P ) has the weak Dixmier property.
By the same argument as in the beginning of the proof of Corollary 4.3, we may assume that M M M has separable predual, see also [32 Theorems D and E(6)] . For the remainder of this section, we therefore assume that M M M has separable predual, while P P P need only be countably decomposable. Fix a normal faithful state ϕ \phi ϕ on M M M , and let ρ = ϕ ∘ E \rho=\phi\circ E ρ = ϕ ∘ E . We use the compatible realizations
c ϕ ( M ) ⊂ c ρ ( P ) = P ⋊ σ ρ R . c_\phi(M) \subset c_\rho(P)=P\rtimes_{\sigma_\rho}\mathbb{R}. c ϕ ( M ) ⊂ c ρ ( P ) = P ⋊ σ ρ R .
We write ϕ i t \phi^{it} ϕ i t for the canonical implementing unitaries in c ϕ ( M ) c_\phi(M) c ϕ ( M ) . Let A = c ρ ( P ) \mathcal{A}=c_\rho(P) A = c ρ ( P ) , and let ψ A = ρ ∘ T \psi_{\mathcal{A}}=\rho\circ T ψ A = ρ ∘ T be the dual weight of ρ \rho ρ , where T : A + → P ^ + T:\mathcal{A}_+\to\widehat{P}_+ T : A + → P + is the canonical operator-valued weight. We use the notation h = d ϕ A d τ A h=\frac{d\phi_{\mathcal{A}}}{d\tau_{\mathcal{A}}} h = d τ A d ϕ A and H = log h H=\log h H = log h from Subsection 2.2, where τ A \tau_{\mathcal{A}} τ A is the canonical trace on A \mathcal{A} A . Here λ t = h i t = ϕ i t \lambda_t=h^{it}=\phi^{it} λ t = h i t = ϕ i t . The modular action σ t ψ A = Ad ( λ t ) \sigma_t^{\psi_{\mathcal{A}}}=\operatorname{Ad}(\lambda_t) σ t ψ A = Ad ( λ t ) restricts to σ t ϕ \sigma_t^\phi σ t ϕ on M M M . Take the ( M , ϕ ) (M,\phi) ( M , ϕ ) -hyperstate φ \varphi φ and the inner ucp coefficients ( y k ) (y_k) ( y k ) from Example 3.9 , and define
Q ( x ) = ∑ k y k x y k ∗ ( x ∈ A ) . (27) Q(x)=\sum_k y_kxy_k^* \qquad(x\in\mathcal{A}). \tag*{(27)} Q ( x ) = k ∑ y k x y k ∗ ( x ∈ A ) . (27)
To prove that M ⊂ A M\subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property, we only need to prove that d : = Q ( a a ∗ ) − a a ∗ = 0 d:=Q(aa^*)-aa^*=0 d := Q ( a a ∗ ) − a a ∗ = 0 for every a ∈ Har ( Q , A ) a\in\operatorname{Har}(Q,\mathcal{A}) a ∈ Har ( Q , A ) . As in the proof of Proposition 3.13 and Lemma 4.5 , followed by Cesàro averaging, we have
Q ( a a ∗ ) = a a ∗ ( a ∈ Har ( Q , A ) ) ⟹ Har ( Q , A ) = M ′ ∩ A ⟹ M ⊂ A has the inner UCP-Dixmier property. \begin{aligned}
Q(aa^*)&=aa^* \qquad(a\in\operatorname{Har}(Q,\mathcal{A}))\\
&\Longrightarrow\operatorname{Har}(Q,\mathcal{A})=M'\cap\mathcal{A}\\
&\Longrightarrow M\subset\mathcal{A}\text{ has the inner UCP-Dixmier property.}
\end{aligned} Q ( a a ∗ ) = a a ∗ ( a ∈ Har ( Q , A )) ⟹ Har ( Q , A ) = M ′ ∩ A ⟹ M ⊂ A has the inner UCP-Dixmier property.
Fix a ∈ Har ( Q , A ) a\in\operatorname{Har}(Q,\mathcal{A}) a ∈ Har ( Q , A ) . By Kadison’s inequality, d ≥ 0 d\ge0 d ≥ 0 . Lemma 2.2 reduces d = 0 d=0 d = 0 to proving that every x ∈ A + x\in\mathcal{A}_+ x ∈ A + with 0 ≤ x ≤ d 0\le x\le d 0 ≤ x ≤ d and ψ A ( x ) < ∞ \psi_{\mathcal{A}}(x)<\infty ψ A ( x ) < ∞ must satisfy x = 0 x=0 x = 0 . Such an x x x satisfies
∑ k = 0 n − 1 Q k ( x ) ≤ ∑ k = 0 n − 1 Q k ( d ) = Q n ( a a ∗ ) − a a ∗ ≤ C 1 ( n ≥ 1 ) , \sum_{k=0}^{n-1}Q^k(x)\le\sum_{k=0}^{n-1}Q^k(d)=Q^n(aa^*)-aa^*\le C1 \qquad(n\ge1), k = 0 ∑ n − 1 Q k ( x ) ≤ k = 0 ∑ n − 1 Q k ( d ) = Q n ( a a ∗ ) − a a ∗ ≤ C 1 ( n ≥ 1 ) ,
with C = ∥ a ∥ 2 C=\lVert a\rVert^2 C = ∥ a ∥ 2 .
To explain how this bound and ψ A ( x ) < ∞ \psi_{\mathcal{A}}(x)<\infty ψ A ( x ) < ∞ imply x = 0 x=0 x = 0 , we first present a commutative model, which is not used in the proof of Theorem A. This model illustrates the classical cutoff argument for recurrence of symmetric Markov operators, see [11 Section 1.6, Theorem 1.6.3] . In this model, L ∞ ( R , d t ) L^\infty(\mathbb{R},dt) L ∞ ( R , d t ) , equipped with the weight induced by Lebesgue measure, plays the role of ( A , ψ A ) (\mathcal{A},\psi_{\mathcal{A}}) ( A , ψ A ) , and a symmetric measure ν ∈ Prob ( R ) \nu\in\operatorname{Prob}(\mathbb{R}) ν ∈ Prob ( R ) plays the role of the symmetric nrsg hyperstate φ \varphi φ .
Proposition 5.1. Let ν ∈ Prob ( R ) \nu\in\operatorname{Prob}(\mathbb{R}) ν ∈ Prob ( R ) be symmetric, and assume that m 2 = ∫ R s 2 d ν ( s ) < ∞ m_2=\int_{\mathbb{R}}s^2\,d\nu(s)<\infty m 2 = ∫ R s 2 d ν ( s ) < ∞ . Define the Markov operator Q Q Q on L ∞ ( R , d t ) L^\infty(\mathbb{R},dt) L ∞ ( R , d t ) by
Q ( b ) ( t ) = ∫ R b ( t + s ) d ν ( s ) ( b ∈ L ∞ ( R , d t ) ) . Q(b)(t)=\int_{\mathbb{R}}b(t+s)\,d\nu(s) \qquad(b\in L^\infty(\mathbb{R},dt)). Q ( b ) ( t ) = ∫ R b ( t + s ) d ν ( s ) ( b ∈ L ∞ ( R , d t )) . Let X ∈ L ∞ ( R , d t ) + X\in L^\infty(\mathbb{R},dt)_+ X ∈ L ∞ ( R , d t ) + with ∫ R X ( t ) d t < ∞ \int_{\mathbb{R}}X(t)\,dt<\infty ∫ R X ( t ) d t < ∞ . If there exists 0 ≤ C < ∞ 0\le C<\infty 0 ≤ C < ∞ such that
∑ k = 0 n − 1 Q k ( X ) ≤ C 1 ( n ≥ 1 ) , (28) \sum_{k=0}^{n-1}Q^k(X)\le C1 \qquad(n\ge1), \tag*{(28)} k = 0 ∑ n − 1 Q k ( X ) ≤ C 1 ( n ≥ 1 ) , (28) then X = 0 X=0 X = 0 .
Proof. We first prove that Q Q Q preserves the integral and is a self-adjoint contraction on L 2 ( R ) L^2(\mathbb{R}) L 2 ( R ) (cf. Lemma 5.3). By Tonelli’s theorem and translation invariance, we have ∫ R Q ( b ) d t = ∫ R b d t \int_{\mathbb{R}} Q(b)\,dt=\int_{\mathbb{R}}b\,dt ∫ R Q ( b ) d t = ∫ R b d t for b ∈ L 1 ( R ) + b\in L^1(\mathbb{R})_+ b ∈ L 1 ( R ) + . By Kadison’s inequality, we then have ∥ Q ( b ) ∥ 2 ≤ ∥ b ∥ 2 \lVert Q(b)\rVert_2\leq\lVert b\rVert_2 ∥ Q ( b ) ∥ 2 ≤ ∥ b ∥ 2 for b ∈ L 2 ∩ L ∞ ( R ) b\in L^2\cap L^\infty(\mathbb{R}) b ∈ L 2 ∩ L ∞ ( R ) , and hence on L 2 ( R ) L^2(\mathbb{R}) L 2 ( R ) by density. Since ν \nu ν is symmetric, Q Q Q is self-adjoint, so 1 − Q ≥ 0 1-Q\geq0 1 − Q ≥ 0 .
We may assume that ∫ R X d t > 0 \int_{\mathbb{R}}X\,dt>0 ∫ R X d t > 0 . Take a real f ∈ C ∞ ( R ) f\in C^\infty(\mathbb{R}) f ∈ C ∞ ( R ) with 0 ≤ f ≤ 1 0\leq f\leq1 0 ≤ f ≤ 1 , equal to 1 1 1 near 0 0 0 , even and nonincreasing in ∣ t ∣ \lvert t\rvert ∣ t ∣ . Let F R ( t ) = f ( t / R ) F_R(t)=f(t/R) F R ( t ) = f ( t / R ) for R ≥ 1 R\geq1 R ≥ 1 . Then F R ↗ 1 F_R\nearrow1 F R ↗ 1 pointwise (R → ∞ R\to\infty R → ∞ ) and ∫ R F R d t < ∞ \int_{\mathbb{R}}F_R\,dt<\infty ∫ R F R d t < ∞ .
We next prove the partial-sum estimate (cf. Lemma 5.4 )
∫ R X d t ≤ lim inf R → ∞ C ∥ ( 1 − Q ) 1 / 2 F R ∥ 2 2 . (29) \int_{\mathbb{R}}X\,dt\leq\liminf_{R\to\infty}C\lVert(1-Q)^{1/2}F_R\rVert_2^2. \tag*{(29)} ∫ R X d t ≤ R → ∞ lim inf C ∥( 1 − Q ) 1/2 F R ∥ 2 2 . (29)
Let a n = ∑ k = 0 n − 1 Q k ( X ) a_n=\sum_{k=0}^{n-1}Q^k(X) a n = ∑ k = 0 n − 1 Q k ( X ) . Then 0 ≤ a n ≤ C 1 0\leq a_n\leq C1 0 ≤ a n ≤ C 1 , ∫ R a n d t = n ∫ R X d t \int_{\mathbb{R}}a_n\,dt=n\int_{\mathbb{R}}X\,dt ∫ R a n d t = n ∫ R X d t , and a n ∈ L 1 ∩ L ∞ ( R ) ⊂ L 2 ( R ) a_n\in L^1\cap L^\infty(\mathbb{R})\subset L^2(\mathbb{R}) a n ∈ L 1 ∩ L ∞ ( R ) ⊂ L 2 ( R ) . Since ( 1 − Q ) a n = X − Q n ( X ) (1-Q)a_n=X-Q^n(X) ( 1 − Q ) a n = X − Q n ( X ) ,
∥ ( 1 − Q ) 1 / 2 a n ∥ 2 2 = ∫ R a n ( X − Q n ( X ) ) d t ≤ C ∫ R X d t . \lVert(1-Q)^{1/2}a_n\rVert_2^2=\int_{\mathbb{R}}a_n(X-Q^n(X))\,dt\leq C\int_{\mathbb{R}}X\,dt. ∥( 1 − Q ) 1/2 a n ∥ 2 2 = ∫ R a n ( X − Q n ( X )) d t ≤ C ∫ R X d t .
For fixed R R R , by the Cauchy–Schwarz inequality we therefore have
∣ ∫ R F R ( X − Q n ( X ) ) d t ∣ 2 ≤ C ( ∫ R X d t ) ∥ ( 1 − Q ) 1 / 2 F R ∥ 2 2 . (30) \left\lvert\int_{\mathbb{R}}F_R(X-Q^n(X))\,dt\right\rvert^2\leq C\left(\int_{\mathbb{R}}X\,dt\right)\lVert(1-Q)^{1/2}F_R\rVert_2^2. \tag*{(30)} ∫ R F R ( X − Q n ( X )) d t 2 ≤ C ( ∫ R X d t ) ∥( 1 − Q ) 1/2 F R ∥ 2 2 . (30)
By (28) and monotone convergence,
∑ n = 0 ∞ ∫ R F R Q n ( X ) d t ≤ C ∫ R F R d t < ∞ . \sum_{n=0}^{\infty}\int_{\mathbb{R}}F_RQ^n(X)\,dt\leq C\int_{\mathbb{R}}F_R\,dt<\infty. n = 0 ∑ ∞ ∫ R F R Q n ( X ) d t ≤ C ∫ R F R d t < ∞.
Hence ∫ R F R Q n ( X ) d t → 0 \int_{\mathbb{R}}F_RQ^n(X)\,dt\to0 ∫ R F R Q n ( X ) d t → 0 (n → ∞ n\to\infty n → ∞ ), so by (30) we have
( ∫ R F R X d t ) 2 ≤ C ( ∫ R X d t ) ∥ ( 1 − Q ) 1 / 2 F R ∥ 2 2 . \left(\int_{\mathbb{R}}F_RX\,dt\right)^2\leq C\left(\int_{\mathbb{R}}X\,dt\right)\lVert(1-Q)^{1/2}F_R\rVert_2^2. ( ∫ R F R X d t ) 2 ≤ C ( ∫ R X d t ) ∥( 1 − Q ) 1/2 F R ∥ 2 2 .
By dominated convergence as R → ∞ R\to\infty R → ∞ and division by ∫ R X d t > 0 \int_{\mathbb{R}}X\,dt>0 ∫ R X d t > 0 , we have (29) .
We now prove the cutoff estimate (cf. Lemma 5.5 ). Expanding the squared difference, for b ∈ L 2 ( R ) b\in L^2(\mathbb{R}) b ∈ L 2 ( R ) we have
∥ ( 1 − Q ) 1 / 2 b ∥ 2 2 = 1 2 ∫ R ∥ b ( ⋅ + s ) − b ∥ 2 2 d ν ( s ) . (31) \lVert(1-Q)^{1/2}b\rVert_2^2=\frac{1}{2}\int_{\mathbb{R}}\lVert b(\mathord{\cdot}+s)-b\rVert_2^2\,d\nu(s). \tag*{(31)} ∥( 1 − Q ) 1/2 b ∥ 2 2 = 2 1 ∫ R ∥ b ( ⋅ + s ) − b ∥ 2 2 d ν ( s ) . (31)
By the fundamental theorem of calculus and the Cauchy–Schwarz inequality, we have
∥ F R ( ⋅ + s ) − F R ∥ 2 ≤ ∣ s ∣ ∥ F R ′ ∥ 2 = ∣ s ∣ R ∥ f ′ ∥ 2 ( s ∈ R ) . (32) \lVert F_R(\mathord{\cdot}+s)-F_R\rVert_2\leq\lvert s\rvert\lVert F_R'\rVert_2=\frac{\lvert s\rvert}{\sqrt{R}}\lVert f'\rVert_2\qquad(s\in\mathbb{R}). \tag*{(32)} ∥ F R ( ⋅ + s ) − F R ∥ 2 ≤ ∣ s ∣ ∥ F R ′ ∥ 2 = R ∣ s ∣ ∥ f ′ ∥ 2 ( s ∈ R ) . (32)
Recall that m 2 = ∫ R s 2 d ν ( s ) < ∞ m_2=\int_{\mathbb{R}}s^2\,d\nu(s)<\infty m 2 = ∫ R s 2 d ν ( s ) < ∞ . By (31) and (32) , we have
∥ ( 1 − Q ) 1 / 2 F R ∥ 2 2 ≤ ∥ f ′ ∥ 2 2 2 R ∫ R s 2 d ν ( s ) = m 2 ∥ f ′ ∥ 2 2 2 R . (33) \lVert(1-Q)^{1/2}F_R\rVert_2^2\leq\frac{\lVert f'\rVert_2^2}{2R}\int_{\mathbb{R}}s^2\,d\nu(s)=\frac{m_2\lVert f'\rVert_2^2}{2R}. \tag*{(33)} ∥( 1 − Q ) 1/2 F R ∥ 2 2 ≤ 2 R ∥ f ′ ∥ 2 2 ∫ R s 2 d ν ( s ) = 2 R m 2 ∥ f ′ ∥ 2 2 . (33)
Hence ∥ ( 1 − Q ) 1 / 2 F R ∥ 2 2 → 0 \lVert(1-Q)^{1/2}F_R\rVert_2^2\to0 ∥( 1 − Q ) 1/2 F R ∥ 2 2 → 0 (R → ∞ R\to\infty R → ∞ ), so by (29) we have ∫ R X d t = 0 \int_{\mathbb{R}}X\,dt=0 ∫ R X d t = 0 and X = 0 X=0 X = 0 . □
Remark 5.2. Symmetry of Q Q Q is essential in Proposition 5.1. Indeed, let Q 0 ( f ) ( t ) = f ( t − 1 ) Q_0(f)(t)=f(t-1) Q 0 ( f ) ( t ) = f ( t − 1 ) be the nonsymmetric translation on L ∞ ( R ) L^\infty(\mathbb{R}) L ∞ ( R ) , and let X = 1 [ 0 , 1 ) X=1_{[0,1)} X = 1 [ 0 , 1 ) . Then X ≠ 0 X\neq0 X = 0 , but
∑ n = 0 ∞ Q 0 n ( X ) = 1 [ 0 , ∞ ) ≤ 1. \sum_{n=0}^{\infty}Q_0^n(X)=1_{[0,\infty)}\leq1. n = 0 ∑ ∞ Q 0 n ( X ) = 1 [ 0 , ∞ ) ≤ 1.
This is why our proof of Theorem A uses the symmetric nrsg hyperstate constructed in Example 3.9, whereas the arguments in Section 4 and later in Section 6 work with an arbitrary nrsg hyperstate.
We now return to Q : A → A Q:\mathcal{A}\to\mathcal{A} Q : A → A defined by (27) and to x ∈ A + x\in\mathcal{A}_{+} x ∈ A + with ψ A ( x ) < ∞ \psi_{\mathcal{A}}(x)<\infty ψ A ( x ) < ∞ and ∑ k = 0 n − 1 Q k ( x ) ≤ C 1 \sum_{k=0}^{n-1}Q^{k}(x)\le C1 ∑ k = 0 n − 1 Q k ( x ) ≤ C 1 (n ≥ 1 n\ge1 n ≥ 1 ). The proof of x = 0 x=0 x = 0 follows the three steps in the proof of Proposition 5.1 . We first prove in (34) that Q Q Q preserves ψ A \psi_{\mathcal{A}} ψ A . By the symmetry (13) of the coefficients ( y k ) (y_k) ( y k ) , we then construct in Lemma 5.3 a self-adjoint contraction S S S on L 2 ( A , τ A ) L^{2}(\mathcal{A},\tau_{\mathcal{A}}) L 2 ( A , τ A ) , which plays the role of Q Q Q on L 2 ( R ) L^{2}(\mathbb{R}) L 2 ( R ) . Lemmas 5.4 and 5.5 are the analogues of (29) and (33) , respectively. Together they prove that ψ A ( x ) = 0 \psi_{\mathcal{A}}(x)=0 ψ A ( x ) = 0 . By faithfulness of ψ A \psi_{\mathcal{A}} ψ A , we have x = 0 x=0 x = 0 .
Hence d = 0 d=0 d = 0 for every a ∈ Har ( Q , A ) a\in\operatorname{Har}(Q,\mathcal{A}) a ∈ Har ( Q , A ) , and M ⊂ A M\subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property. By Theorem 4.1, we then have the weak Dixmier property when M M M is properly infinite. For finite M M M , we take ϕ \phi ϕ tracial so that the Cesàro means are already unitary averages. The general case follows by decomposing M M M into its finite and properly infinite parts.
Since ϕ ∘ Q = ϕ \phi\circ Q=\phi ϕ ∘ Q = ϕ , by normality and M M M -bimodularity of E E E and T T T , we have
ψ A ∘ Q = ϕ ∘ E ∘ T ∘ Q = ϕ ∘ Q ∘ E ∘ T = ϕ ∘ E ∘ T = ψ A , (34) \psi_{\mathcal{A}}\circ Q=\phi\circ E\circ T\circ Q=\phi\circ Q\circ E\circ T=\phi\circ E\circ T=\psi_{\mathcal{A}}, \tag*{(34)} ψ A ∘ Q = ϕ ∘ E ∘ T ∘ Q = ϕ ∘ Q ∘ E ∘ T = ϕ ∘ E ∘ T = ψ A , (34)
where the compositions are understood on the extended positive cone.
Define
m ψ A = span { a ∈ A + ∣ ψ A ( a ) < ∞ } , ξ ( a ) = h 1 / 4 a h 1 / 4 ( a ∈ m ψ A ) . m_{\psi_{\mathcal{A}}}=\operatorname{span}\{a\in\mathcal{A}_{+}\mid\psi_{\mathcal{A}}(a)<\infty\},\qquad\xi(a)=h^{1/4}ah^{1/4}\quad(a\in m_{\psi_{\mathcal{A}}}). m ψ A = span { a ∈ A + ∣ ψ A ( a ) < ∞ } , ξ ( a ) = h 1/4 a h 1/4 ( a ∈ m ψ A ) .
On L 2 ( A , τ A ) L^{2}(\mathcal{A},\tau_{\mathcal{A}}) L 2 ( A , τ A ) we use the inner product ⟨ X , Y ⟩ 2 = τ A ( Y ∗ X ) \langle X,Y\rangle_{2}=\tau_{\mathcal{A}}(Y^{*}X) ⟨ X , Y ⟩ 2 = τ A ( Y ∗ X ) , linear in the first variable.
Define S : D → D S:D\to D S : D → D , where D = { ξ ( a ) ∣ a ∈ m ψ A } D=\{\xi(a)\mid a\in m_{\psi_{\mathcal{A}}}\} D = { ξ ( a ) ∣ a ∈ m ψ A } , by
S ξ ( a ) = ξ ( Q ( a ) ) ( a ∈ m ψ A ) . (35) S\xi(a)=\xi(Q(a))\qquad(a\in m_{\psi_{\mathcal{A}}}). \tag*{(35)} S ξ ( a ) = ξ ( Q ( a )) ( a ∈ m ψ A ) . (35)
This is well-defined since ξ \xi ξ is injective and Q ( m ψ A ) ⊂ m ψ A Q(m_{\psi_{\mathcal{A}}})\subset m_{\psi_{\mathcal{A}}} Q ( m ψ A ) ⊂ m ψ A by (34) .
Lemma 5.3. For every a ∈ m ψ A a\in m_{\psi_{\mathcal{A}}} a ∈ m ψ A , we have ξ ( a ) ∈ L 2 ( A , τ A ) \xi(a)\in L^{2}(\mathcal{A},\tau_{\mathcal{A}}) ξ ( a ) ∈ L 2 ( A , τ A ) . Let x ∈ A + x\in\mathcal{A}_{+} x ∈ A + satisfy ψ A ( x ) < ∞ \psi_{\mathcal{A}}(x)<\infty ψ A ( x ) < ∞ . Then
ω x ( a ) = τ A ( h 1 / 2 x h 1 / 2 a ) ( a ∈ A ) \omega_x(a)=\tau_{\mathcal{A}}(h^{1/2}xh^{1/2}a)\qquad(a\in\mathcal{A}) ω x ( a ) = τ A ( h 1/2 x h 1/2 a ) ( a ∈ A ) defines a finite normal positive functional on A \mathcal{A} A , with ω x ( 1 ) = ψ A ( x ) \omega_x(1)=\psi_{\mathcal{A}}(x) ω x ( 1 ) = ψ A ( x ) , and
⟨ ξ ( a ) , ξ ( x ) ⟩ 2 = ω x ( a ) ( a ∈ m ψ A ) . (36) \langle\xi(a),\xi(x)\rangle_{2}=\omega_x(a)\qquad(a\in m_{\psi_{\mathcal{A}}}). \tag*{(36)} ⟨ ξ ( a ) , ξ ( x ) ⟩ 2 = ω x ( a ) ( a ∈ m ψ A ) . (36) Moreover, S S S extends to a self-adjoint contraction on L 2 ( A , τ A ) L^{2}(\mathcal{A},\tau_{\mathcal{A}}) L 2 ( A , τ A ) .
Proof. For a ∈ m ψ A a\in m_{\psi_{\mathcal{A}}} a ∈ m ψ A , let
ι ( a ) = h 1 / 2 a h 1 / 2 ∈ L 1 ( A , τ A ) . \iota(a)=h^{1/2}ah^{1/2}\in L^{1}(\mathcal{A},\tau_{\mathcal{A}}). ι ( a ) = h 1/2 a h 1/2 ∈ L 1 ( A , τ A ) .
We first prove the assertions about ξ \xi ξ and ω x \omega_x ω x . Both a h 1 / 2 ah^{1/2} a h 1/2 and h 1 / 2 a h^{1/2}a h 1/2 a belong to L 2 ( A , τ A ) L^{2}(\mathcal{A},\tau_{\mathcal{A}}) L 2 ( A , τ A ) . By traciality and the Cauchy–Schwarz inequality, we have
∥ ξ ( a ) ∥ 2 2 = τ A ( h 1 / 4 a ∗ h 1 / 2 a h 1 / 4 ) = τ A ( ( a h 1 / 2 ) ∗ ( h 1 / 2 a ) ) ≤ ∥ a h 1 / 2 ∥ 2 ∥ h 1 / 2 a ∥ 2 = τ A ( h 1 / 2 a ∗ a h 1 / 2 ) 1 / 2 τ A ( h 1 / 2 a a ∗ h 1 / 2 ) 1 / 2 = ψ A ( a ∗ a ) 1 / 2 ψ A ( a a ∗ ) 1 / 2 . (37) \begin{aligned}
\|\xi(a)\|_{2}^{2}
&=\tau_{\mathcal{A}}(h^{1/4}a^{*}h^{1/2}ah^{1/4})\\
&=\tau_{\mathcal{A}}((ah^{1/2})^{*}(h^{1/2}a))\\
&\le\|ah^{1/2}\|_{2}\|h^{1/2}a\|_{2}\\
&=\tau_{\mathcal{A}}(h^{1/2}a^{*}ah^{1/2})^{1/2}\tau_{\mathcal{A}}(h^{1/2}aa^{*}h^{1/2})^{1/2}\\
&=\psi_{\mathcal{A}}(a^{*}a)^{1/2}\psi_{\mathcal{A}}(aa^{*})^{1/2}. \tag*{(37)}
\end{aligned} ∥ ξ ( a ) ∥ 2 2 = τ A ( h 1/4 a ∗ h 1/2 a h 1/4 ) = τ A (( a h 1/2 ) ∗ ( h 1/2 a )) ≤ ∥ a h 1/2 ∥ 2 ∥ h 1/2 a ∥ 2 = τ A ( h 1/2 a ∗ a h 1/2 ) 1/2 τ A ( h 1/2 a a ∗ h 1/2 ) 1/2 = ψ A ( a ∗ a ) 1/2 ψ A ( a a ∗ ) 1/2 . (37)
We first make this calculation for a n = e n a e n a_n=e_na e_n a n = e n a e n , where e n = 1 [ 1 / n , n ] ( h ) e_n=\mathbf{1}_{[1/n,n]}(h) e n = 1 [ 1/ n , n ] ( h ) . The convergences a n h 1 / 2 → a h 1 / 2 a_nh^{1/2}\to ah^{1/2} a n h 1/2 → a h 1/2 and h 1 / 2 a n → h 1 / 2 a h^{1/2}a_n\to h^{1/2}a h 1/2 a n → h 1/2 a in L 2 L^{2} L 2 justify passage to the limit.
Since ι ( x ) ∈ L 1 ( A , τ A ) + \iota(x)\in L^{1}(\mathcal{A},\tau_{\mathcal{A}})_{+} ι ( x ) ∈ L 1 ( A , τ A ) + and τ A ( ι ( x ) ) = ψ A ( x ) \tau_{\mathcal{A}}(\iota(x))=\psi_{\mathcal{A}}(x) τ A ( ι ( x )) = ψ A ( x ) , ω x \omega_x ω x is a finite normal positive functional with ω x ( 1 ) = ψ A ( x ) \omega_x(1)=\psi_{\mathcal{A}}(x) ω x ( 1 ) = ψ A ( x ) . To prove (36) , let
e n = 1 [ 1 / n , n ] ( h ) , x n = e n x e n , a n = e n a e n . e_n=\mathbf{1}_{[1/n,n]}(h),\qquad x_n=e_nxe_n,\qquad a_n=e_nae_n. e n = 1 [ 1/ n , n ] ( h ) , x n = e n x e n , a n = e n a e n .
By (6) , the projections e n e_n e n have finite trace. By traciality in e n A e n e_n\mathcal{A}e_n e n A e n , we have
⟨ ξ ( a n ) , ξ ( x n ) ⟩ 2 = τ A ( h 1 / 2 x n h 1 / 2 a n ) . \langle\xi(a_n),\xi(x_n)\rangle_{2}=\tau_{\mathcal{A}}(h^{1/2}x_nh^{1/2}a_n). ⟨ ξ ( a n ) , ξ ( x n ) ⟩ 2 = τ A ( h 1/2 x n h 1/2 a n ) .
Since e n h = h e n e_nh=he_n e n h = h e n and e n ↗ 1 e_n \nearrow1 e n ↗ 1 strongly,
ι ( x n ) = e n ι ( x ) e n ⟶ ι ( x ) in L 1 , ξ ( x n ) = e n ξ ( x ) e n ⟶ ξ ( x ) in L 2 , ξ ( a n ) ⟶ ξ ( a ) in L 2 . \iota(x_n)=e_n\iota(x)e_n \longrightarrow\iota(x)\quad\text{in }L^1,\qquad\xi(x_n)=e_n\xi(x)e_n \longrightarrow\xi(x)\quad\text{in }L^2,\qquad\xi(a_n)\longrightarrow\xi(a)\quad\text{in }L^2. ι ( x n ) = e n ι ( x ) e n ⟶ ι ( x ) in L 1 , ξ ( x n ) = e n ξ ( x ) e n ⟶ ξ ( x ) in L 2 , ξ ( a n ) ⟶ ξ ( a ) in L 2 .
Moreover,
sup n ∥ a n ∥ ≤ ∥ a ∥ , a n ⟶ a ∗ -strongly . \sup_n\|a_n\|\leq\|a\|,\qquad a_n\longrightarrow a\quad*\text{-strongly}. n sup ∥ a n ∥ ≤ ∥ a ∥ , a n ⟶ a ∗ -strongly .
Passing to the limit proves (36) .
We next prove that S S S extends to a self-adjoint contraction. Identifying A ∗ A_* A ∗ with L 1 ( A , τ A ) L^1(A,\tau_A) L 1 ( A , τ A ) , we first show that the preadjoint of Q Q Q satisfies
Q ∗ ι ( a ) = ι ( Q ( a ) ) ( a ∈ m ψ A ) . (38) Q_*\iota(a)=\iota(Q(a))\qquad(a\in\mathfrak m_{\psi_A}). \tag*{(38)} Q ∗ ι ( a ) = ι ( Q ( a )) ( a ∈ m ψ A ) . (38)
Let v k = σ i / 2 φ ( y k ∗ ) v_k=\sigma_{i/2}^{\varphi}(y_k^*) v k = σ i /2 φ ( y k ∗ ) . By Example 3.9 , ( v k ) (v_k) ( v k ) is a permutation of ( y k ) (y_k) ( y k ) . We have
y k ∗ h 1 / 2 = h 1 / 2 v k , h 1 / 2 y k = v k ∗ h 1 / 2 . (39) y_k^*h^{1/2}=h^{1/2}v_k,\qquad h^{1/2}y_k=v_k^*h^{1/2}. \tag*{(39)} y k ∗ h 1/2 = h 1/2 v k , h 1/2 y k = v k ∗ h 1/2 . (39)
Formula (39) follows first on compact spectral vectors for H H H , and then for the closed measurable products. For a ∈ A + a\in\mathcal A_+ a ∈ A + with ψ A ( a ) < ∞ \psi_A(a)<\infty ψ A ( a ) < ∞ , we have
y k ∗ ι ( a ) y k = ι ( v k a v k ∗ ) . y_k^*\iota(a)y_k=\iota(v_kav_k^*). y k ∗ ι ( a ) y k = ι ( v k a v k ∗ ) .
By (39) , the positive partial sums ∑ k ≤ n y k ∗ ι ( a ) y k \sum_{k\leq n}y_k^*\iota(a)y_k ∑ k ≤ n y k ∗ ι ( a ) y k converge in L 1 L^1 L 1 , since
τ A ( ∑ k ≤ n y k ∗ ι ( a ) y k ) = τ A ( ι ( a ) ∑ k ≤ n y k y k ∗ ) ↗ ψ A ( a ) . \tau_A\left(\sum_{k\leq n}y_k^*\iota(a)y_k\right)=\tau_A\left(\iota(a)\sum_{k\leq n}y_ky_k^*\right)\nearrow\psi_A(a). τ A ( k ≤ n ∑ y k ∗ ι ( a ) y k ) = τ A ( ι ( a ) k ≤ n ∑ y k y k ∗ ) ↗ ψ A ( a ) .
By monotone convergence of the corresponding quadratic forms, we have
Q ∗ ι ( a ) = ∑ k y k ∗ ι ( a ) y k = ι ( ∑ k v k a v k ∗ ) = ι ( Q ( a ) ) . Q_*\iota(a)=\sum_k y_k^*\iota(a)y_k=\iota\left(\sum_kv_kav_k^*\right)=\iota(Q(a)). Q ∗ ι ( a ) = k ∑ y k ∗ ι ( a ) y k = ι ( k ∑ v k a v k ∗ ) = ι ( Q ( a )) .
This proves (38) for positive a a a , and the general case follows by linearity.
By (38) and (36) , for a , b ∈ m ψ A ∩ A + a,b\in\mathfrak m_{\psi_A}\cap\mathcal A_+ a , b ∈ m ψ A ∩ A + ,
⟨ S ξ ( a ) , ξ ( b ) ⟩ 2 = τ A ( ι ( b ) Q ( a ) ) = τ A ( Q ∗ ι ( b ) a ) = τ A ( ι ( Q ( b ) ) a ) = ⟨ ξ ( a ) , S ξ ( b ) ⟩ 2 . \langle S\xi(a),\xi(b)\rangle_2=\tau_A(\iota(b)Q(a))=\tau_A(Q_*\iota(b)a)=\tau_A(\iota(Q(b))a)=\langle\xi(a),S\xi(b)\rangle_2. ⟨ S ξ ( a ) , ξ ( b ) ⟩ 2 = τ A ( ι ( b ) Q ( a )) = τ A ( Q ∗ ι ( b ) a ) = τ A ( ι ( Q ( b )) a ) = ⟨ ξ ( a ) , S ξ ( b ) ⟩ 2 .
By linearity, we have ⟨ S v , w ⟩ 2 = ⟨ v , S w ⟩ 2 \langle Sv,w\rangle_2=\langle v,Sw\rangle_2 ⟨ S v , w ⟩ 2 = ⟨ v , S w ⟩ 2 for v , w ∈ D v,w\in D v , w ∈ D . Also S ( D ) ⊂ D S(D)\subset D S ( D ) ⊂ D by (34) . For a ∈ m ψ A a\in\mathfrak m_{\psi_A} a ∈ m ψ A , by (35) , (37) , Kadison’s inequality for Q n Q^n Q n , and ψ A ∘ Q n = ψ A \psi_A\circ Q^n=\psi_A ψ A ∘ Q n = ψ A , we have
∥ S n ξ ( a ) ∥ 2 2 = ∥ ξ ( Q n ( a ) ) ∥ 2 2 = ∥ h 1 / 4 Q n ( a ) h 1 / 4 ∥ 2 2 = τ A ( h 1 / 2 Q n ( a ) ∗ h 1 / 2 Q n ( a ) ) ≤ ψ A ( Q n ( a ) ∗ Q n ( a ) ) 1 / 2 ψ A ( Q n ( a ) Q n ( a ) ∗ ) 1 / 2 ≤ ψ A ( Q n ( a ∗ a ) ) 1 / 2 ψ A ( Q n ( a a ∗ ) ) 1 / 2 = ψ A ( a ∗ a ) 1 / 2 ψ A ( a a ∗ ) 1 / 2 . \begin{aligned}
\|S^n\xi(a)\|_2^2&=\|\xi(Q^n(a))\|_2^2\\
&=\|h^{1/4}Q^n(a)h^{1/4}\|_2^2\\
&=\tau_A(h^{1/2}Q^n(a)^*h^{1/2}Q^n(a))\\
&\leq\psi_A(Q^n(a)^*Q^n(a))^{1/2}\psi_A(Q^n(a)Q^n(a)^*)^{1/2}\\
&\leq\psi_A(Q^n(a^*a))^{1/2}\psi_A(Q^n(aa^*))^{1/2}\\
&=\psi_A(a^*a)^{1/2}\psi_A(aa^*)^{1/2}.
\end{aligned} ∥ S n ξ ( a ) ∥ 2 2 = ∥ ξ ( Q n ( a )) ∥ 2 2 = ∥ h 1/4 Q n ( a ) h 1/4 ∥ 2 2 = τ A ( h 1/2 Q n ( a ) ∗ h 1/2 Q n ( a )) ≤ ψ A ( Q n ( a ) ∗ Q n ( a ) ) 1/2 ψ A ( Q n ( a ) Q n ( a ) ∗ ) 1/2 ≤ ψ A ( Q n ( a ∗ a ) ) 1/2 ψ A ( Q n ( a a ∗ ) ) 1/2 = ψ A ( a ∗ a ) 1/2 ψ A ( a a ∗ ) 1/2 .
Hence
C v : = sup m ≥ 0 ∥ S m v ∥ 2 < ∞ ( v ∈ D ) . C_v:=\sup_{m\geq0}\|S^mv\|_2<\infty\qquad(v\in D). C v := m ≥ 0 sup ∥ S m v ∥ 2 < ∞ ( v ∈ D ) .
For v ∈ D v\in D v ∈ D , by symmetry, we have
∥ S v ∥ 2 2 = ⟨ S 2 v , v ⟩ 2 ≤ ∥ S 2 v ∥ 2 ∥ v ∥ 2 . (40) \|Sv\|_2^2=\langle S^2v,v\rangle_2\leq\|S^2v\|_2\|v\|_2. \tag*{(40)} ∥ S v ∥ 2 2 = ⟨ S 2 v , v ⟩ 2 ≤ ∥ S 2 v ∥ 2 ∥ v ∥ 2 . (40)
Iterating (40) , we have
∥ S v ∥ 2 ≤ ∥ S 2 n v ∥ 2 1 / 2 n ∥ v ∥ 2 1 − 1 / 2 n ≤ C v 1 / 2 n ∥ v ∥ 2 1 − 1 / 2 n ⟶ ∥ v ∥ 2 . \|Sv\|_2\leq\|S^{2^n}v\|_2^{1/2^n}\|v\|_2^{1-1/2^n}\leq C_v^{1/2^n}\|v\|_2^{1-1/2^n}\longrightarrow\|v\|_2. ∥ S v ∥ 2 ≤ ∥ S 2 n v ∥ 2 1/ 2 n ∥ v ∥ 2 1 − 1/ 2 n ≤ C v 1/ 2 n ∥ v ∥ 2 1 − 1/ 2 n ⟶ ∥ v ∥ 2 .
It remains to check that D D D is dense. With e n = 1 [ 1 / n , n ] ( h ) e_n=\mathbf{1}_{[1/n,n]}(h) e n = 1 [ 1/ n , n ] ( h ) , both h h h and h − 1 h^{-1} h − 1 are bounded on e n e_n e n , and τ A ( e n ) < ∞ \tau_A(e_n)<\infty τ A ( e n ) < ∞ . Every bounded b ∈ e n A e n b\in e_nAe_n b ∈ e n A e n has the form
b = ξ ( h − 1 / 4 b h − 1 / 4 ) , h − 1 / 4 b h − 1 / 4 ∈ e n A e n ⊂ m ψ A . b=\xi(h^{-1/4}bh^{-1/4}),\qquad h^{-1/4}bh^{-1/4}\in e_nAe_n\subset m_{\psi_A}. b = ξ ( h − 1/4 b h − 1/4 ) , h − 1/4 b h − 1/4 ∈ e n A e n ⊂ m ψ A .
These elements are dense in L 2 ( A , τ A ) L^2(A,\tau_A) L 2 ( A , τ A ) . Therefore S S S extends to a self-adjoint contraction. □
Let f ∈ C ∞ ( R ) f\in C^\infty(\mathbb{R}) f ∈ C ∞ ( R ) be real with 0 ≤ f ≤ 1 0\le f\le1 0 ≤ f ≤ 1 , equal to 1 near 0, even and nonincreasing in ∣ s ∣ |s| ∣ s ∣ . Define H R = f ( H / R ) H_R=f(H/R) H R = f ( H / R ) , R ≥ 1 R\ge1 R ≥ 1 . Then H R ↗ 1 H_R\nearrow1 H R ↗ 1 strongly as R → ∞ R\to\infty R → ∞ , and by (5) , we have
ψ A ( H R ) = R ∫ R f ( s ) d s < ∞ . \psi_A(H_R)=R\int_{\mathbb{R}}f(s)\,\mathrm{d}s<\infty. ψ A ( H R ) = R ∫ R f ( s ) d s < ∞.
Lemma 5.4. Let x ∈ A + x\in A_+ x ∈ A + satisfy ψ A ( x ) < ∞ \psi_A(x)<\infty ψ A ( x ) < ∞ . Assume that there is a constant 0 ≤ C < ∞ 0\le C<\infty 0 ≤ C < ∞ such that
∑ k = 0 n − 1 Q k ( x ) ≤ C 1 ( n ≥ 1 ) . (41) \sum_{k=0}^{n-1}\mathcal{Q}^k(x)\le C1\qquad(n\ge1). \tag*{(41)} k = 0 ∑ n − 1 Q k ( x ) ≤ C 1 ( n ≥ 1 ) . (41) Then
ψ A ( x ) ≤ lim inf R → ∞ C ∥ ( 1 − S ) 1 / 2 ξ ( H R ) ∥ 2 2 . (42) \psi_A(x)\le\liminf_{R\to\infty}C\left\|(1-S)^{1/2}\xi(H_R)\right\|_2^2. \tag*{(42)} ψ A ( x ) ≤ R → ∞ lim inf C ( 1 − S ) 1/2 ξ ( H R ) 2 2 . (42)
Proof. The case ψ A ( x ) = 0 \psi_A(x)=0 ψ A ( x ) = 0 is immediate. Assume that ψ A ( x ) > 0 \psi_A(x)>0 ψ A ( x ) > 0 . Let a n = ∑ k = 0 n − 1 Q k ( x ) ≤ C 1 a_n=\sum_{k=0}^{n-1}\mathcal{Q}^k(x)\le C1 a n = ∑ k = 0 n − 1 Q k ( x ) ≤ C 1 . By (34) and (35) ,
ψ A ( a n ) = n ψ A ( x ) < ∞ , ( 1 − S ) ξ ( a n ) = ξ ( x − Q n ( x ) ) . \psi_A(a_n)=n\psi_A(x)<\infty,\qquad(1-S)\xi(a_n)=\xi(x-\mathcal{Q}^n(x)). ψ A ( a n ) = n ψ A ( x ) < ∞ , ( 1 − S ) ξ ( a n ) = ξ ( x − Q n ( x )) .
By (36) and positivity,
∥ ( 1 − S ) 1 / 2 ξ ( a n ) ∥ 2 2 = ⟨ ξ ( x − Q n ( x ) ) , ξ ( a n ) ⟩ 2 = ω x ( a n ) − ω a n ( Q n ( x ) ) ≤ ω x ( a n ) ≤ C ω x ( 1 ) = C ψ A ( x ) . (43) \left\|(1-S)^{1/2}\xi(a_n)\right\|_2^2
=\left\langle\xi(x-\mathcal{Q}^n(x)),\xi(a_n)\right\rangle_2
=\omega_x(a_n)-\omega_{a_n}(\mathcal{Q}^n(x))
\le\omega_x(a_n)\le C\omega_x(1)=C\psi_A(x). \tag*{(43)} ( 1 − S ) 1/2 ξ ( a n ) 2 2 = ⟨ ξ ( x − Q n ( x )) , ξ ( a n ) ⟩ 2 = ω x ( a n ) − ω a n ( Q n ( x )) ≤ ω x ( a n ) ≤ C ω x ( 1 ) = C ψ A ( x ) . (43)
Fix R ≥ 1 R\ge1 R ≥ 1 . By (36) and positivity, ω x ( H R ) = ω H R ( x ) ‾ = ω H R ( x ) \omega_x(H_R)=\overline{\omega_{H_R}(x)}=\omega_{H_R}(x) ω x ( H R ) = ω H R ( x ) = ω H R ( x ) . Since a n − Q ( a n ) = x − Q n ( x ) a_n-\mathcal{Q}(a_n)=x-\mathcal{Q}^n(x) a n − Q ( a n ) = x − Q n ( x ) and 1 − S ≥ 0 1-S\ge0 1 − S ≥ 0 , by (35) and (43) , we have
∣ ω x ( H R ) − ω H R ( Q n ( x ) ) ∣ 2 = ∣ ω H R ( x − Q n ( x ) ) ∣ 2 = ∣ ⟨ ξ ( x − Q n ( x ) ) , ξ ( H R ) ⟩ 2 ∣ 2 = ∣ ⟨ ( 1 − S ) ξ ( a n ) , ξ ( H R ) ⟩ 2 ∣ 2 = ∣ ⟨ ( 1 − S ) 1 / 2 ξ ( a n ) , ( 1 − S ) 1 / 2 ξ ( H R ) ⟩ 2 ∣ 2 ≤ ∥ ( 1 − S ) 1 / 2 ξ ( a n ) ∥ 2 2 ∥ ( 1 − S ) 1 / 2 ξ ( H R ) ∥ 2 2 ≤ C ψ A ( x ) ∥ ( 1 − S ) 1 / 2 ξ ( H R ) ∥ 2 2 . (44) \begin{aligned}
\left|\omega_x(H_R)-\omega_{H_R}(\mathcal{Q}^n(x))\right|^2
&=\left|\omega_{H_R}(x-\mathcal{Q}^n(x))\right|^2\\
&=\left|\left\langle\xi(x-\mathcal{Q}^n(x)),\xi(H_R)\right\rangle_2\right|^2\\
&=\left|\left\langle(1-S)\xi(a_n),\xi(H_R)\right\rangle_2\right|^2\\
&=\left|\left\langle(1-S)^{1/2}\xi(a_n),(1-S)^{1/2}\xi(H_R)\right\rangle_2\right|^2\\
&\le\left\|(1-S)^{1/2}\xi(a_n)\right\|_2^2\left\|(1-S)^{1/2}\xi(H_R)\right\|_2^2\\
&\le C\psi_A(x)\left\|(1-S)^{1/2}\xi(H_R)\right\|_2^2. \tag*{(44)}
\end{aligned} ∣ ω x ( H R ) − ω H R ( Q n ( x )) ∣ 2 = ∣ ω H R ( x − Q n ( x )) ∣ 2 = ∣ ⟨ ξ ( x − Q n ( x )) , ξ ( H R ) ⟩ 2 ∣ 2 = ∣ ⟨ ( 1 − S ) ξ ( a n ) , ξ ( H R ) ⟩ 2 ∣ 2 = ⟨ ( 1 − S ) 1/2 ξ ( a n ) , ( 1 − S ) 1/2 ξ ( H R ) ⟩ 2 2 ≤ ( 1 − S ) 1/2 ξ ( a n ) 2 2 ( 1 − S ) 1/2 ξ ( H R ) 2 2 ≤ C ψ A ( x ) ( 1 − S ) 1/2 ξ ( H R ) 2 2 . (44)
By (41) , for every N ≥ 1 N\ge1 N ≥ 1 ,
ω H R ( Q n ( x ) ) = ω H R ( ∑ n = 0 N − 1 Q n ( x ) ) ≤ C ω H R ( 1 ) = C ψ A ( H R ) < ∞ . \omega_{H_R}\left(\mathcal{Q}^n(x)\right)
=\omega_{H_R}\left(\sum_{n=0}^{N-1}\mathcal{Q}^n(x)\right)
\le C\omega_{H_R}(1)=C\psi_A(H_R)<\infty. ω H R ( Q n ( x ) ) = ω H R ( n = 0 ∑ N − 1 Q n ( x ) ) ≤ C ω H R ( 1 ) = C ψ A ( H R ) < ∞.
Letting N → ∞ N\to\infty N → ∞ , we have
∑ n = 0 ∞ ω H R ( Q n ( x ) ) ≤ C ψ A ( H R ) < ∞ , \sum_{n=0}^{\infty}\omega_{H_R}\left(\mathcal{Q}^n(x)\right)\le C\psi_A(H_R)<\infty, n = 0 ∑ ∞ ω H R ( Q n ( x ) ) ≤ C ψ A ( H R ) < ∞ ,
so ω H R ( Q n ( x ) ) → 0 \omega_{H_R}(\mathcal{Q}^n(x))\to0 ω H R ( Q n ( x )) → 0 as n → ∞ n\to\infty n → ∞ . Letting n → ∞ n\to\infty n → ∞ in (44) , we have
∣ ω x ( H R ) ∣ 2 ≤ C ψ A ( x ) ∥ ( 1 − S ) 1 / 2 ξ ( H R ) ∥ 2 2 ( R ≥ 1 ) . (45) |\omega_x(H_R)|^2\le C\psi_A(x)\left\|(1-S)^{1/2}\xi(H_R)\right\|_2^2\qquad(R\ge1). \tag*{(45)} ∣ ω x ( H R ) ∣ 2 ≤ C ψ A ( x ) ( 1 − S ) 1/2 ξ ( H R ) 2 2 ( R ≥ 1 ) . (45)
Since H R ↗ 1 H_R \nearrow1 H R ↗ 1 strongly as R → ∞ R \to\infty R → ∞ , by normality of ω x \omega_x ω x and (45) , we have
ψ A ( x ) 2 = ∣ ω x ( 1 ) ∣ 2 = lim R → ∞ ∣ ω x ( H R ) ∣ 2 ≤ lim inf R → ∞ C ψ A ( x ) ∥ ( 1 − S ) 1 / 2 ξ ( H R ) ∥ 2 2 . \psi_A(x)^2 = |\omega_x(1)|^2 = \lim_{R\to\infty} |\omega_x(H_R)|^2 \le\liminf_{R\to\infty} C\psi_A(x)\|(1-S)^{1/2}\xi(H_R)\|_2^2. ψ A ( x ) 2 = ∣ ω x ( 1 ) ∣ 2 = R → ∞ lim ∣ ω x ( H R ) ∣ 2 ≤ R → ∞ lim inf C ψ A ( x ) ∥ ( 1 − S ) 1/2 ξ ( H R ) ∥ 2 2 .
Dividing by ψ A ( x ) \psi_A(x) ψ A ( x ) proves (42) . □
Let ν ∈ Prob ( R ) \nu\in\operatorname{Prob}(\mathbb{R}) ν ∈ Prob ( R ) be the spectral distribution of log Δ ϕ \log\Delta_\phi log Δ ϕ under φ \varphi φ , i.e., ν ( E ) = φ ( 1 E ( log Δ ϕ ) ) \nu(E) = \varphi(\mathbf{1}_E(\log\Delta_\phi)) ν ( E ) = φ ( 1 E ( log Δ ϕ )) for every Borel set E ⊂ R E \subset\mathbb{R} E ⊂ R . Define
m 2 : = ∫ R s 2 d ν ( s ) = φ ( ( log Δ ϕ ) 2 ) . m_2 := \int_{\mathbb{R}} s^2\,d\nu(s) = \varphi((\log\Delta_\phi)^2). m 2 := ∫ R s 2 d ν ( s ) = φ (( log Δ ϕ ) 2 ) .
Lemma 5.5. We have m 2 < ∞ m_2 < \infty m 2 < ∞ and, for every R ≥ 1 R \ge1 R ≥ 1 ,
∥ ( 1 − S ) 1 / 2 ξ ( H R ) ∥ 2 2 ≤ m 2 ∥ f ′ ∥ 2 2 2 R . (46) \|(1-S)^{1/2}\xi(H_R)\|_2^2 \le\frac{m_2\|f'\|_2^2}{2R}. \tag*{(46)} ∥ ( 1 − S ) 1/2 ξ ( H R ) ∥ 2 2 ≤ 2 R m 2 ∥ f ′ ∥ 2 2 . (46) Moreover, if x ∈ A + x \in A_+ x ∈ A + satisfies ψ A ( x ) < ∞ \psi_A(x) < \infty ψ A ( x ) < ∞ and ∑ k = 0 n − 1 Q k ( x ) ≤ C 1 \sum_{k=0}^{n-1} Q^k(x) \le C1 ∑ k = 0 n − 1 Q k ( x ) ≤ C 1 for some 0 ≤ C < ∞ 0 \le C < \infty 0 ≤ C < ∞ and every n ≥ 1 n \ge1 n ≥ 1 , then x = 0 x = 0 x = 0 .
Proof. We first prove that m 2 < ∞ m_2 < \infty m 2 < ∞ . By (17) , the Tomita identity, and Example 3.9(ii), we have
φ ( Δ ϕ ) = ∑ k ⟨ S ϕ y k ∗ ξ ϕ , S ϕ y k ∗ ξ ϕ ⟩ = ∑ k ∥ y k ξ ϕ ∥ 2 ≤ ∑ k ∥ y k ∥ ∞ 2 < ∞ . (47) \varphi(\Delta_\phi) = \sum_k \langle S_\phi y_k^*\xi_\phi,S_\phi y_k^*\xi_\phi\rangle= \sum_k \|y_k\xi_\phi\|^2 \le\sum_k \|y_k\|_\infty^2 < \infty. \tag*{(47)} φ ( Δ ϕ ) = k ∑ ⟨ S ϕ y k ∗ ξ ϕ , S ϕ y k ∗ ξ ϕ ⟩ = k ∑ ∥ y k ξ ϕ ∥ 2 ≤ k ∑ ∥ y k ∥ ∞ 2 < ∞. (47)
Since J Δ ϕ J = Δ ϕ − 1 J\Delta_\phi J = \Delta_\phi^{-1} J Δ ϕ J = Δ ϕ − 1 , by the symmetry in (12) , we have φ ( Δ ϕ − 1 ) = φ ( Δ ϕ ) < ∞ \varphi(\Delta_\phi^{-1}) = \varphi(\Delta_\phi) < \infty φ ( Δ ϕ − 1 ) = φ ( Δ ϕ ) < ∞ . By (47) and ( log t ) 2 ≤ t + t − 1 − 2 (\log t)^2 \le t + t^{-1} - 2 ( log t ) 2 ≤ t + t − 1 − 2 (t > 0 t > 0 t > 0 ), we have m 2 = φ ( ( log Δ ϕ ) 2 ) < ∞ m_2 = \varphi((\log\Delta_\phi)^2) < \infty m 2 = φ (( log Δ ϕ ) 2 ) < ∞ .
Let g ∈ C ∞ ( R ) g \in C^\infty(\mathbb{R}) g ∈ C ∞ ( R ) be real. By (5) , g ( H ) ∈ m ψ A g(H) \in\mathfrak{m}_{\psi_A} g ( H ) ∈ m ψ A , and g ( H ) ∈ A ψ A g(H) \in A_{\psi_A} g ( H ) ∈ A ψ A . Expanding the commutators, we have
∑ k [ y k , g ( H ) ] [ y k , g ( H ) ] ∗ = Q ( g ( H ) 2 ) − Q ( g ( H ) ) g ( H ) − g ( H ) Q ( g ( H ) ) + g ( H ) 2 . (48) \sum_k [y_k,g(H)][y_k,g(H)]^* = Q(g(H)^2) - Q(g(H))g(H) - g(H)Q(g(H)) + g(H)^2. \tag*{(48)} k ∑ [ y k , g ( H )] [ y k , g ( H ) ] ∗ = Q ( g ( H ) 2 ) − Q ( g ( H )) g ( H ) − g ( H ) Q ( g ( H )) + g ( H ) 2 . (48)
By Kadison’s inequality, (34) , and (5) ,
ψ A ( Q ( g ( H ) ) 2 ) ≤ ψ A ( Q ( g ( H ) 2 ) ) = ψ A ( g ( H ) 2 ) = ∫ R ∣ g ( s ) ∣ 2 d s < ∞ . \psi_A(Q(g(H))^2) \le\psi_A(Q(g(H)^2)) = \psi_A(g(H)^2) = \int_{\mathbb{R}} |g(s)|^2\,ds < \infty. ψ A ( Q ( g ( H ) ) 2 ) ≤ ψ A ( Q ( g ( H ) 2 )) = ψ A ( g ( H ) 2 ) = ∫ R ∣ g ( s ) ∣ 2 d s < ∞.
Hence all four terms on the right of (48) belong to m ψ A \mathfrak{m}_{\psi_A} m ψ A . Applying ψ A \psi_A ψ A and using the ψ A \psi_A ψ A -centralizer property of g ( H ) g(H) g ( H ) , the linear extension of (36) to self-adjoint elements, and (35) , we have
∑ k ψ A ( [ y k , g ( H ) ] [ y k , g ( H ) ] ∗ ) = 2 ( ψ A ( g ( H ) 2 ) − ψ A ( g ( H ) Q ( g ( H ) ) ) ) = 2 ( ⟨ ξ ( g ( H ) ) , ξ ( g ( H ) ) ⟩ 2 − ⟨ S ξ ( g ( H ) ) , ξ ( g ( H ) ) ⟩ 2 ) = 2 ∥ ( 1 − S ) 1 / 2 ξ ( g ( H ) ) ∥ 2 2 . (49) \begin{aligned}
\sum_k \psi_A([y_k,g(H)][y_k,g(H)]^*) &= 2(\psi_A(g(H)^2)-\psi_A(g(H)Q(g(H))))\\
&= 2(\langle\xi(g(H)),\xi(g(H))\rangle_2-\langle S\xi(g(H)),\xi(g(H))\rangle_2)\\
&= 2\|(1-S)^{1/2}\xi(g(H))\|_2^2. \tag*{(49)}
\end{aligned} k ∑ ψ A ([ y k , g ( H )] [ y k , g ( H ) ] ∗ ) = 2 ( ψ A ( g ( H ) 2 ) − ψ A ( g ( H ) Q ( g ( H )))) = 2 (⟨ ξ ( g ( H )) , ξ ( g ( H )) ⟩ 2 − ⟨ S ξ ( g ( H )) , ξ ( g ( H )) ⟩ 2 ) = 2∥ ( 1 − S ) 1/2 ξ ( g ( H )) ∥ 2 2 . (49)
We compute the left-hand side by proving
∑ k ψ A ( [ y k , g ( H ) ] [ y k , g ( H ) ] ∗ ) = ∫ R ∫ R ∣ g ( r + s ) − g ( r ) ∣ 2 d r d ν ( s ) . (50) \sum_k \psi_A([y_k,g(H)][y_k,g(H)]^*) = \int_{\mathbb{R}}\int_{\mathbb{R}} |g(r+s)-g(r)|^2\,dr\,d\nu(s). \tag*{(50)} k ∑ ψ A ([ y k , g ( H )] [ y k , g ( H ) ] ∗ ) = ∫ R ∫ R ∣ g ( r + s ) − g ( r ) ∣ 2 d r d ν ( s ) . (50)
We use the unitary Fourier transform
g ^ ( t ) = ( 2 π ) − 1 / 2 ∫ R e − i t r g ( r ) d r . \widehat{g}(t) = (2\pi)^{-1/2}\int_{\mathbb{R}} e^{-itr}g(r)\,dr. g ( t ) = ( 2 π ) − 1/2 ∫ R e − i t r g ( r ) d r .
By the crossed-product relation, we have
[ y k , g ( H ) ] = ( 2 π ) − 1 / 2 ∫ R g ^ ( t ) ( y k − σ t ϕ ( y k ) ) λ t d t . (51) [y_k,g(H)] = (2\pi)^{-1/2}\int_{\mathbb{R}} \widehat{g}(t)(y_k-\sigma_t^\phi(y_k))\lambda_t\,dt. \tag*{(51)} [ y k , g ( H )] = ( 2 π ) − 1/2 ∫ R g ( t ) ( y k − σ t ϕ ( y k )) λ t d t . (51)
With the normalization in (5) , for X = ( 2 π ) − 1 / 2 ∫ R a t λ t d t = ( 2 π ) − 1 / 2 ∫ R a t e i t H d t X=(2\pi)^{-1/2}\int_{\mathbb{R}}a_t\lambda_t\,dt=(2\pi)^{-1/2}\int_{\mathbb{R}}a_te^{itH}\,dt X = ( 2 π ) − 1/2 ∫ R a t λ t d t = ( 2 π ) − 1/2 ∫ R a t e i t H d t , we have
ψ A ( X ∗ X ) = ∫ R ρ ( a t ∗ a t ) d t , ψ A ( X X ∗ ) = ∫ R ρ ( a t a t ∗ ) d t . (52) \psi_A(X^*X)=\int_{\mathbb{R}}\rho(a_t^*a_t)\,dt,\qquad\psi_A(XX^*)=\int_{\mathbb{R}}\rho(a_ta_t^*)\,dt. \tag*{(52)} ψ A ( X ∗ X ) = ∫ R ρ ( a t ∗ a t ) d t , ψ A ( X X ∗ ) = ∫ R ρ ( a t a t ∗ ) d t . (52)
The identities in (52) hold first on the crossed-product Hilbert algebra and extend by its L 2 L^2 L 2 isometry, see [16 ] . In the first identity in (52) , the zeroth coefficient is ∫ R ρ ( σ − t φ ( a t ∗ a t ) ) d t \int_{\mathbb{R}}\rho\bigl(\sigma_{-t}^{\varphi}(a_t^*a_t)\bigr)\,dt ∫ R ρ ( σ − t φ ( a t ∗ a t ) ) d t , and the modular invariance of ρ \rho ρ removes this twist.
By φ = ∑ k ⟨ y k ∗ ξ φ , y k ∗ ξ φ ⟩ \varphi=\sum_k\langle y_k^*\xi_\varphi,y_k^*\xi_\varphi\rangle φ = ∑ k ⟨ y k ∗ ξ φ , y k ∗ ξ φ ⟩ and the definition of ν \nu ν as the spectral distribution of log Δ φ \log\Delta_\varphi log Δ φ under φ \varphi φ , we have
∑ k φ ( ( y k − σ t φ ( y k ) ) ( y k − σ t φ ( y k ) ) ∗ ) = ∑ k ∥ ( 1 − Δ φ i t ) y k ∗ ξ φ ∥ 2 = ∑ k ⟨ ( 1 − Δ φ − i t ) ( 1 − Δ φ i t ) y k ∗ ξ φ , y k ∗ ξ φ ⟩ = φ ( ( 1 − Δ φ − i t ) ( 1 − Δ φ i t ) ) = ∫ R ( 1 − e − i t s ) ( 1 − e i t s ) d ν ( s ) = ∫ R ∣ 1 − e i t s ∣ 2 d ν ( s ) . \begin{aligned}
\sum_k\varphi\bigl((y_k-\sigma_t^\varphi(y_k))(y_k-\sigma_t^\varphi(y_k))^*\bigr)
&=\sum_k\lVert(1-\Delta_\varphi^{it})y_k^*\xi_\varphi\rVert^2\\
&=\sum_k\langle(1-\Delta_\varphi^{-it})(1-\Delta_\varphi^{it})y_k^*\xi_\varphi,y_k^*\xi_\varphi\rangle\\
&=\varphi\bigl((1-\Delta_\varphi^{-it})(1-\Delta_\varphi^{it})\bigr)\\
&=\int_{\mathbb{R}}(1-e^{-its})(1-e^{its})\,d\nu(s)\\
&=\int_{\mathbb{R}}|1-e^{its}|^2\,d\nu(s).
\end{aligned} k ∑ φ ( ( y k − σ t φ ( y k )) ( y k − σ t φ ( y k ) ) ∗ ) = k ∑ ∥( 1 − Δ φ i t ) y k ∗ ξ φ ∥ 2 = k ∑ ⟨( 1 − Δ φ − i t ) ( 1 − Δ φ i t ) y k ∗ ξ φ , y k ∗ ξ φ ⟩ = φ ( ( 1 − Δ φ − i t ) ( 1 − Δ φ i t ) ) = ∫ R ( 1 − e − i t s ) ( 1 − e i t s ) d ν ( s ) = ∫ R ∣1 − e i t s ∣ 2 d ν ( s ) .
By (51) , (52) , and scalar Plancherel,
∑ k ψ A ( [ y k , g ( H ) ] [ y k , g ( H ) ] ∗ ) = ∫ R ∣ g ^ ( t ) ∣ 2 ∫ R ∣ 1 − e i t s ∣ 2 d ν ( s ) d t = ∫ R ∫ R ∣ g ^ ( t ) ∣ 2 ∣ 1 − e i t s ∣ 2 d t d ν ( s ) = ∫ R ∫ R ∣ g ( r + s ) − g ( r ) ∣ 2 d r d ν ( s ) . \begin{aligned}
\sum_k\psi_A([y_k,g(H)][y_k,g(H)]^*)
&=\int_{\mathbb{R}}|\widehat{g}(t)|^2\int_{\mathbb{R}}|1-e^{its}|^2\,d\nu(s)\,dt\\
&=\int_{\mathbb{R}}\int_{\mathbb{R}}|\widehat{g}(t)|^2|1-e^{its}|^2\,dt\,d\nu(s)\\
&=\int_{\mathbb{R}}\int_{\mathbb{R}}|g(r+s)-g(r)|^2\,dr\,d\nu(s).
\end{aligned} k ∑ ψ A ([ y k , g ( H )] [ y k , g ( H ) ] ∗ ) = ∫ R ∣ g ( t ) ∣ 2 ∫ R ∣1 − e i t s ∣ 2 d ν ( s ) d t = ∫ R ∫ R ∣ g ( t ) ∣ 2 ∣1 − e i t s ∣ 2 d t d ν ( s ) = ∫ R ∫ R ∣ g ( r + s ) − g ( r ) ∣ 2 d r d ν ( s ) .
This proves (50) . Now take g ( r ) = f ( r / R ) g(r)=f(r/R) g ( r ) = f ( r / R ) , so g ( H ) = H R g(H)=H_R g ( H ) = H R . Since ∥ g ( ⋅ + s ) − g ∥ 2 ≤ ∣ s ∣ ∥ g ′ ∥ 2 = ∣ s ∣ R ∥ f ′ ∥ 2 \lVert g(\cdot+s)-g\rVert_2\leq|s|\lVert g'\rVert_2=\frac{|s|}{\sqrt{R}}\lVert f'\rVert_2 ∥ g ( ⋅ + s ) − g ∥ 2 ≤ ∣ s ∣ ∥ g ′ ∥ 2 = R ∣ s ∣ ∥ f ′ ∥ 2 , by (50) , we have
∑ k ψ A ( [ y k , H R ] [ y k , H R ] ∗ ) ≤ m 2 ∥ f ′ ∥ 2 2 R . (53) \sum_k\psi_A([y_k,H_R][y_k,H_R]^*)\leq\frac{m_2\lVert f'\rVert_2^2}{R}. \tag*{(53)} k ∑ ψ A ([ y k , H R ] [ y k , H R ] ∗ ) ≤ R m 2 ∥ f ′ ∥ 2 2 . (53)
Combining (49) with g ( H ) = H R g(H)=H_R g ( H ) = H R and (53) proves (46) .
Now let x x x satisfy the hypotheses of Lemma 5.4. By (42) and (46) ,
0 ≤ ψ A ( x ) ≤ lim inf R → ∞ C m 2 ∥ f ′ ∥ 2 2 2 R = 0. 0\leq\psi_A(x)\leq\liminf_{R\to\infty}\frac{Cm_2\lVert f'\rVert_2^2}{2R}=0. 0 ≤ ψ A ( x ) ≤ R → ∞ lim inf 2 R C m 2 ∥ f ′ ∥ 2 2 = 0.
Hence ψ A ( x ) = 0 \psi_A(x)=0 ψ A ( x ) = 0 , and by faithfulness of ψ A \psi_A ψ A , we have x = 0 x=0 x = 0 . □
We can now identify Har ( Q , A ) \operatorname{Har}(Q,\mathcal{A}) Har ( Q , A ) .
Proposition 5.6. The fixed-point space of Q Q Q is
Har ( Q , A ) = M ′ ∩ A . \operatorname{Har}(Q,\mathcal{A})=M'\cap\mathcal{A}. Har ( Q , A ) = M ′ ∩ A .
Proof. We only need to prove that Har ( Q , A ) ⊂ M ′ ∩ A \operatorname{Har}(Q,\mathcal{A})\subset M'\cap\mathcal{A} Har ( Q , A ) ⊂ M ′ ∩ A . Let a ∈ Har ( Q , A ) a\in\operatorname{Har}(Q,\mathcal{A}) a ∈ Har ( Q , A ) . By Kadison’s inequality,
d : = Q ( a a ∗ ) − a a ∗ = Q ( a a ∗ ) − Q ( a ) Q ( a ) ∗ ≥ 0. d:=Q(aa^*)-aa^*=Q(aa^*)-Q(a)Q(a)^*\geq0. d := Q ( a a ∗ ) − a a ∗ = Q ( a a ∗ ) − Q ( a ) Q ( a ) ∗ ≥ 0.
We have
∑ k = 0 n − 1 Q k ( d ) = Q n ( a a ∗ ) − a a ∗ ≤ ∥ a ∥ 1 2 ( n ≥ 1 ) . \sum_{k=0}^{n-1}Q^k(d)=Q^n(aa^*)-aa^*\leq\lVert a\rVert_1^2\qquad(n\geq1). k = 0 ∑ n − 1 Q k ( d ) = Q n ( a a ∗ ) − a a ∗ ≤ ∥ a ∥ 1 2 ( n ≥ 1 ) .
Suppose that d ≠ 0 d \ne0 d = 0 . By Lemma 2.2, take 0 ≠ x ≤ d 0 \ne x \le d 0 = x ≤ d with ψ A ( x ) < ∞ \psi_A(x) < \infty ψ A ( x ) < ∞ . Then
∑ k = 0 n − 1 Q k ( x ) ≤ ∑ k = 0 n − 1 Q k ( d ) ≤ ∥ a ∥ 1 2 ( n ≥ 1 ) . \sum_{k=0}^{n-1} \mathcal{Q}^{k}(x) \le\sum_{k=0}^{n-1} \mathcal{Q}^{k}(d) \le\lVert a\rVert_1^2 \qquad(n \ge1). k = 0 ∑ n − 1 Q k ( x ) ≤ k = 0 ∑ n − 1 Q k ( d ) ≤ ∥ a ∥ 1 2 ( n ≥ 1 ) .
By Lemma 5.5 , we have x = 0 x=0 x = 0 , a contradiction. Hence Q ( a a ∗ ) = a a ∗ Q(aa^*)=aa^* Q ( a a ∗ ) = a a ∗ . Applying the same argument to a ∗ ∈ Har ( Q , A ) a^* \in\operatorname{Har}(Q,\mathcal{A}) a ∗ ∈ Har ( Q , A ) , we have Q ( a ∗ a ) = a ∗ a Q(a^*a)=a^*a Q ( a ∗ a ) = a ∗ a , so a a a belongs to the multiplicative domain of Q Q Q . Expanding as in (21) , we have
∑ k ( y k a − a y k ) ( y k a − a y k ) ∗ = 0. \sum_k (y_k a-a y_k)(y_k a-a y_k)^*=0. k ∑ ( y k a − a y k ) ( y k a − a y k ) ∗ = 0.
Hence [ a , y k ] = 0 [a,y_k]=0 [ a , y k ] = 0 for every k k k . Since C [ ( y k ) ] ‾ w = M \overline{\mathbb{C}[(y_k)]}^{\,w}=M C [( y k )] w = M , a ∈ M ′ ∩ A a\in M'\cap\mathcal{A} a ∈ M ′ ∩ A .
Marrakchi proved the relative bicentralizer conjecture, in particular, when M M M is amenable, has no type I I I 1 \mathrm{III}_1 III 1 direct summand, is a tensor product of two type I I I 1 \mathrm{III}_1 III 1 von Neumann algebras, or admits a Cartan subalgebra [32 Theorem E(1)–(4)] . For completeness, we give a new proof covering these cases as well. We now finish the proof by applying Theorem 4.1.
Proof of Theorem A. As in the beginning of the proof of Corollary 4.3, we may assume that M M M has separable predual.
Suppose first that M M M is either finite or properly infinite. Fix a normal faithful state ϕ \phi ϕ on M M M , taking ϕ \phi ϕ tracial when M M M is finite. Let Q Q Q be the map in (27) on A = c ϕ ∘ E ( P ) \mathcal{A}=c_{\phi\circ E}(P) A = c ϕ ∘ E ( P ) , with coefficients from Example 3.9. For x ∈ A x\in\mathcal{A} x ∈ A , define
C n ( x ) = 1 n ∑ k = 0 n − 1 Q k ( x ) , K x = conv w { u x u ∗ ∣ u ∈ U ( M ) } . C_n(x)=\frac{1}{n}\sum_{k=0}^{n-1}Q^k(x), \qquad K_x=\operatorname{conv}^{\,w}\{uxu^*\mid u\in\mathcal{U}(M)\}. C n ( x ) = n 1 k = 0 ∑ n − 1 Q k ( x ) , K x = conv w { ux u ∗ ∣ u ∈ U ( M )} .
Since
∥ Q ( C n ( x ) ) − C n ( x ) ∥ ≤ 2 ∥ x ∥ n , \lVert Q(C_n(x))-C_n(x)\rVert\le\frac{2\lVert x\rVert}{n}, ∥ Q ( C n ( x )) − C n ( x )∥ ≤ n 2 ∥ x ∥ ,
normality of Q Q Q and Proposition 5.6 imply that every ultraweak cluster point of ( C n ( x ) ) (C_n(x)) ( C n ( x )) belongs to M ′ ∩ A M'\cap\mathcal{A} M ′ ∩ A .
If M M M is properly infinite, each C n C_n C n is an M M M -inner ucp map. Hence M ⊂ A M\subset\mathcal{A} M ⊂ A has the inner UCP-Dixmier property, and by Theorem 4.1, we have the weak Dixmier property.
If M M M is finite, by the tracial choice in Example 3.9, we have
y k = λ k u k , Q ( x ) = ∑ k λ k 2 u k x u k ∗ , ∑ k λ k 2 = 1. y_k=\lambda_k u_k,\qquad Q(x)=\sum_k\lambda_k^2u_kxu_k^*,\qquad\sum_k\lambda_k^2=1. y k = λ k u k , Q ( x ) = k ∑ λ k 2 u k x u k ∗ , k ∑ λ k 2 = 1.
Hence Q ( K x ) ⊂ K x Q(K_x)\subset K_x Q ( K x ) ⊂ K x and C n ( x ) ∈ K x C_n(x)\in K_x C n ( x ) ∈ K x for every n n n . Every ultraweak cluster point belongs to K x ∩ ( M ′ ∩ A ) K_x\cap(M'\cap\mathcal{A}) K x ∩ ( M ′ ∩ A ) , which proves the finite case.
For general M M M , write
M = M f i n ⊕ M ∞ M=M_{\mathrm{fin}}\oplus M_{\infty} M = M fin ⊕ M ∞
for its finite and properly infinite parts. We may assume that both summands are nonzero. Let p 1 , p 2 p_1,p_2 p 1 , p 2 be their units and let P i = p i P p i P_i=p_iPp_i P i = p i P p i . There is a normal faithful conditional expectation
F : P → P 1 ⊕ P 2 , F ( a ) = p 1 a p 1 + p 2 a p 2 . F:P\to P_1\oplus P_2,\qquad F(a)=p_1ap_1+p_2ap_2. F : P → P 1 ⊕ P 2 , F ( a ) = p 1 a p 1 + p 2 a p 2 .
The restrictions
E 1 = E ∣ P 1 : P 1 → M f i n , E 2 = E ∣ P 2 : P 2 → M ∞ E_1=E|_{P_1}:P_1\to M_{\mathrm{fin}},\qquad E_2=E|_{P_2}:P_2\to M_{\infty} E 1 = E ∣ P 1 : P 1 → M fin , E 2 = E ∣ P 2 : P 2 → M ∞
are normal faithful conditional expectations.
Take a normal faithful state ϕ \phi ϕ on M M M whose restriction to M f i n M_{\mathrm{fin}} M fin is tracial, and let ρ = ϕ ∘ E \rho=\phi\circ E ρ = ϕ ∘ E , A = c ρ ( P ) \mathcal{A}=c_\rho(P) A = c ρ ( P ) , and ρ i = ρ ∣ P i / ρ ( p i ) \rho_i=\rho|_{P_i}/\rho(p_i) ρ i = ρ ∣ P i / ρ ( p i ) . Since p i ∈ M ϕ p_i\in M_\phi p i ∈ M ϕ , by bimodularity, we have p i ∈ P ρ p_i\in P_\rho p i ∈ P ρ . Hence F F F preserves ρ \rho ρ and extends to a normal faithful conditional expectation
F ~ : A → p 1 A p 1 ⊕ p 2 A p 2 , p i A p i ≅ c ρ i ( P i ) . \widetilde{F}:\mathcal{A}\to p_1\mathcal{A}p_1\oplus p_2\mathcal{A}p_2,\qquad p_i\mathcal{A}p_i\cong c_{\rho_i}(P_i). F : A → p 1 A p 1 ⊕ p 2 A p 2 , p i A p i ≅ c ρ i ( P i ) .
For x ∈ A x \in\mathcal{A} x ∈ A , since p 1 − p 2 ∈ U ( M ) p_1-p_2 \in\mathcal{U}(M) p 1 − p 2 ∈ U ( M ) ,
F ~ ( x ) = p 1 x p 1 + p 2 x p 2 = 1 2 ( x + ( p 1 − p 2 ) x ( p 1 − p 2 ) ) ∈ K x . \widetilde{F}(x)=p_1xp_1+p_2xp_2=\frac{1}{2}\left(x+(p_1-p_2)x(p_1-p_2)\right)\in K_x. F ( x ) = p 1 x p 1 + p 2 x p 2 = 2 1 ( x + ( p 1 − p 2 ) x ( p 1 − p 2 ) ) ∈ K x .
By the finite and properly infinite cases, we have the weak Dixmier property for M f i n ⊂ p 1 A p 1 M_{\mathrm{fin}}\subset p_1\mathcal{A}p_1 M fin ⊂ p 1 A p 1 and M ∞ ⊂ p 2 A p 2 M_\infty\subset p_2\mathcal{A}p_2 M ∞ ⊂ p 2 A p 2 , respectively. As in the proof of Corollary 4.3, taking product nets of the two unitary averaging nets, we have
K x ∩ ( M ′ ∩ A ) ≠ ∅ . K_x\cap(M'\cap\mathcal{A})\ne\varnothing. K x ∩ ( M ′ ∩ A ) = ∅ .
This proves that M ⊂ c ( P ) M\subset c(P) M ⊂ c ( P ) has the weak Dixmier property. Taking P = M P=M P = M , we have the last assertion. □ \square □
Proof of Corollary B. By Theorem A, M ⊂ c ( M ) M\subset c(M) M ⊂ c ( M ) has the weak Dixmier property. Since M M M is of type I I I 1 \mathrm{III}_1 III 1 , by [20 Theorem 3.1] , we have B ( M , ϕ ) = C 1 B(M,\phi)=\mathbb{C}1 B ( M , ϕ ) = C 1 for every normal faithful state ϕ \phi ϕ on M M M . □ \square □
Proof of Corollary C. By Theorem A, we have the weak Dixmier property for M ⊂ c ( P ) M\subset c(P) M ⊂ c ( P ) . Hence by [32 Theorem D] , we have (9) for every normal faithful state ϕ \phi ϕ on M M M . □ \square □
Kadison’s problem on MASA
In this section, we use the generating inner ucp map associated with an nrsg hyperstate to give a new proof of the solution to Kadison’s problem (Theorem G).
Fix an inclusion M ⊂ P M\subset P M ⊂ P of von Neumann algebras with separable preduals and a normal faithful conditional expectation E : P → M E:P\to M E : P → M as in Theorem G .
Following the well-established route of Popa [38 ] , also used in [32 , 33 ] , we only need to construct an increasing sequence of finite-dimensional abelian subalgebras D 1 ⊂ D 2 ⊂ ⋯ ⊂ M D_1\subset D_2\subset\cdots\subset M D 1 ⊂ D 2 ⊂ ⋯ ⊂ M such that ∥ θ ∣ D n ′ ∩ P ∥ → 0 \lVert\theta|_{D_n'\cap P}\rVert\to0 ∥ θ ∣ D n ′ ∩ P ∥ → 0 as n → ∞ n\to\infty n → ∞ for every θ ∈ M ⊥ = { θ ∈ P ∗ ∣ θ ∣ M = 0 } \theta\in M_\perp=\{\theta\in P_* \mid\theta|_M=0\} θ ∈ M ⊥ = { θ ∈ P ∗ ∣ θ ∣ M = 0 } . Then A 0 = ⋁ n D n A_0=\bigvee_nD_n A 0 = ⋁ n D n satisfies ∥ θ ∣ A 0 ′ ∩ P ∥ ≤ ∥ θ ∣ D n ′ ∩ P ∥ → 0 \lVert\theta|_{A_0'\cap P}\rVert\leq\lVert\theta|_{D_n'\cap P}\rVert\to0 ∥ θ ∣ A 0 ′ ∩ P ∥ ≤ ∥ θ ∣ D n ′ ∩ P ∥ → 0 as n → ∞ n\to\infty n → ∞ for every θ ∈ M ⊥ \theta\in M_\perp θ ∈ M ⊥ . Hence, by the Hahn–Banach theorem, A 0 ′ ∩ P ⊂ M A_0'\cap P\subset M A 0 ′ ∩ P ⊂ M . Any MASA A ⊂ M A\subset M A ⊂ M containing A 0 A_0 A 0 then satisfies A ′ ∩ P = A A'\cap P=A A ′ ∩ P = A , which finishes the proof.
Given ε > 0 \varepsilon>0 ε > 0 and θ 1 , … , θ s ∈ M ⊥ \theta_1,\ldots,\theta_s\in M_\perp θ 1 , … , θ s ∈ M ⊥ , we first construct a single finite-dimensional abelian subalgebra D ⊂ M D\subset M D ⊂ M such that ∥ θ j ∣ D ′ ∩ P ∥ < ε \lVert\theta_j|_{D'\cap P}\rVert<\varepsilon ∥ θ j ∣ D ′ ∩ P ∥ < ε for 1 ≤ j ≤ s 1\leq j\leq s 1 ≤ j ≤ s . We will then strengthen this in Proposition 6.7 to construct an increasing sequence D 1 ⊂ D 2 ⊂ ⋯ ⊂ M D_1\subset D_2\subset\cdots\subset M D 1 ⊂ D 2 ⊂ ⋯ ⊂ M with the required convergence.
As mentioned in Subsection 1.3, our main contribution is to use Poisson boundary theory to substantially simplify the construction of D D D in the continuous properly infinite case. This replaces the passage through a hyperfinite subalgebra used by Marrakchi [33 Theorem B] , which in the type I I I 1 \mathrm{III}_1 III 1 case relies on the ergodicity of the relative bicentralizer flow [33 Theorem C] .
Following Popa [40 Definition 2.2] , we call an inclusion M ⊂ P M\subset P M ⊂ P MASA-ergodic if there is an abelian subalgebra A ⊂ M A\subset M A ⊂ M with A ′ ∩ P = A A'\cap P=A A ′ ∩ P = A . The following lemma reduces the proof of Theorem G and the construction of D D D with ∥ θ j ∣ D ′ ∩ P ∥ < ε \lVert\theta_j|_{D'\cap P}\rVert<\varepsilon ∥ θ j ∣ D ′ ∩ P ∥ < ε to the type I, finite, and properly infinite cases.
Lemma 6.1. Let M ⊂ P M\subset P M ⊂ P be as in Theorem G . Then the following hold.
(i) For every nonzero projection e ∈ M e\in M e ∈ M , the restriction E ∣ e P e : e P e → e M e E|_{eP e}:ePe\to eMe E ∣ e P e : e P e → e M e is a normal faithful conditional expectation, and
( e M e ) ′ ∩ e P e = e ( M ′ ∩ P ) = e Z ( M ) = Z ( e M e ) . (54) (eMe)'\cap ePe=e(M'\cap P)=e\mathcal{Z}(M)=\mathcal{Z}(eMe). \tag*{(54)} ( e M e ) ′ ∩ e P e = e ( M ′ ∩ P ) = e Z ( M ) = Z ( e M e ) . (54) Moreover, let ( e k ) ⊂ M (e_k)\subset M ( e k ) ⊂ M be a countable family of mutually orthogonal nonzero projections with ∑ k e k = 1 \sum_k e_k=1 ∑ k e k = 1 . If each inclusion e k M e k ⊂ e k P e k e_kMe_k\subset e_kPe_k e k M e k ⊂ e k P e k is MASA-ergodic, then M ⊂ P M\subset P M ⊂ P is MASA-ergodic.
(ii) For every finite-dimensional abelian subalgebra D ⊂ M D\subset M D ⊂ M ,
( D ′ ∩ M ) ′ ∩ P = D ∨ Z ( M ) ⊂ M . (D'\cap M)'\cap P=D\vee\mathcal{Z}(M)\subset M. ( D ′ ∩ M ) ′ ∩ P = D ∨ Z ( M ) ⊂ M .
(iii) Let e 1 , … , e m ∈ M e_1,\ldots,e_m \in M e 1 , … , e m ∈ M be mutually orthogonal nonzero projections with ∑ i e i = 1 \sum_i e_i = 1 ∑ i e i = 1 , and let D i ⊂ e i M e i D_i \subset e_iMe_i D i ⊂ e i M e i be finite-dimensional abelian subalgebras. Then D = ⨁ i = 1 m D i D = \bigoplus_{i=1}^{m} D_i D = ⨁ i = 1 m D i is a finite-dimensional abelian subalgebra of M M M , and
∥ θ ∣ D ′ ∩ P ∥ ≤ ∑ i = 1 m ∥ θ ∣ D i ′ ∩ e i P e i ∥ ( θ ∈ P ∗ ) . \lVert\theta|_{D'\cap P}\rVert\le\sum_{i=1}^{m}\lVert\theta|_{D_i'\cap e_iPe_i}\rVert\qquad(\theta\in P_*). ∥ θ ∣ D ′ ∩ P ∥ ≤ i = 1 ∑ m ∥ θ ∣ D i ′ ∩ e i P e i ∥ ( θ ∈ P ∗ ) .
Proof. (i) Since E E E is M M M -bimodular, E ( e x e ) = e E ( x ) e E(exe)=eE(x)e E ( e x e ) = e E ( x ) e for x ∈ P x\in P x ∈ P . Hence E ∣ e P e E|_{ePe} E ∣ e P e is a normal faithful conditional expectation onto e M e eMe e M e . By [2 Propositions 4.2.1 and 4.2.2] (see also [26 ] ) and M ′ ∩ P = Z ( M ) M'\cap P=Z(M) M ′ ∩ P = Z ( M ) , we have (54) . For each k k k , take A k ⊂ e k M e k A_k\subset e_kMe_k A k ⊂ e k M e k maximal abelian in e k P e k e_kPe_k e k P e k , and let A = ⨁ k A k ⊂ M A=\bigoplus_k A_k\subset M A = ⨁ k A k ⊂ M . Since e k ∈ A e_k\in A e k ∈ A , for every x ∈ A ′ ∩ P x\in A'\cap P x ∈ A ′ ∩ P we have x = ∑ k e k x e k x=\sum_k e_kxe_k x = ∑ k e k x e k and e k x e k ∈ A k e_kxe_k\in A_k e k x e k ∈ A k for every k k k . Hence A ′ ∩ P = A A'\cap P=A A ′ ∩ P = A .
(ii) Let e 1 , … , e r e_1,\ldots,e_r e 1 , … , e r be the minimal projections of D D D , and let N = D ′ ∩ M N=D'\cap M N = D ′ ∩ M , so that
D = ⨁ i = 1 r C e i , N = ⨁ i = 1 r e i M e i . D=\bigoplus_{i=1}^{r}\mathbb{C}e_i,\qquad N=\bigoplus_{i=1}^{r}e_iMe_i. D = i = 1 ⨁ r C e i , N = i = 1 ⨁ r e i M e i .
If x ∈ N ′ ∩ P x\in N'\cap P x ∈ N ′ ∩ P , then x x x commutes with every e i ∈ N e_i\in N e i ∈ N , and by (54) we have
x = ∑ i = 1 r e i x e i , e i x e i ∈ ( e i M e i ) ′ ∩ e i P e i = e i Z ( M ) . x=\sum_{i=1}^{r}e_ix e_i,\qquad e_ix e_i\in(e_iMe_i)'\cap e_iPe_i=e_iZ(M). x = i = 1 ∑ r e i x e i , e i x e i ∈ ( e i M e i ) ′ ∩ e i P e i = e i Z ( M ) .
Hence N ′ ∩ P ⊂ ⨁ i e i Z ( M ) = D ∨ Z ( M ) N'\cap P\subset\bigoplus_i e_iZ(M)=D\vee Z(M) N ′ ∩ P ⊂ ⨁ i e i Z ( M ) = D ∨ Z ( M ) . The reverse inclusion follows from D ∨ Z ( M ) ⊂ Z ( N ) D\vee Z(M)\subset Z(N) D ∨ Z ( M ) ⊂ Z ( N ) .
(iii) Since e i ∈ D e_i\in D e i ∈ D for every i i i , we have
D ′ ∩ P = ⨁ i = 1 m ( D i ′ ∩ e i P e i ) . D'\cap P=\bigoplus_{i=1}^{m}(D_i'\cap e_iPe_i). D ′ ∩ P = i = 1 ⨁ m ( D i ′ ∩ e i P e i ) .
The estimate follows by the triangle inequality.
By Lemma 6.1(i), we only need to prove Theorem G in the type I and continuous cases. We treat the type I case in the final proof. For continuous M M M , Lemma 6.1(iii) reduces the construction of D D D to the finite and properly infinite cases. We now construct, for any θ 1 , … , θ s ∈ M ⊥ \theta_1,\ldots,\theta_s\in M_\perp θ 1 , … , θ s ∈ M ⊥ and ε > 0 \varepsilon>0 ε > 0 , a finite-dimensional abelian subalgebra D ⊂ M D\subset M D ⊂ M such that ∥ θ j ∣ D ′ ∩ P ∥ < ε \lVert\theta_j|_{D'\cap P}\rVert<\varepsilon ∥ θ j ∣ D ′ ∩ P ∥ < ε for 1 ≤ j ≤ s 1\le j\le s 1 ≤ j ≤ s .
The continuous properly infinite case
We begin with the case where M M M is continuous and properly infinite, i.e., M = M I I ∞ ⊕ M I I I M=M_{II_\infty}\oplus M_{III} M = M I I ∞ ⊕ M I I I . We proceed in two steps.
Using our Poisson boundary tools, we first construct a full matrix algebra F ⊂ M F\subset M F ⊂ M such that ∥ θ j ∣ F ′ ∩ P ∥ < ε / 2 \lVert\theta_j|_{F'\cap P}\rVert<\varepsilon/2 ∥ θ j ∣ F ′ ∩ P ∥ < ε /2 for 1 ≤ j ≤ s 1\le j\le s 1 ≤ j ≤ s (Lemma 6.2 and Proposition 6.3 ).
We then prove the existence of a finite-dimensional abelian subalgebra D ⊂ M D\subset M D ⊂ M such that
∥ θ j ∣ D ′ ∩ P ∥ < ∥ θ j ∣ F ′ ∩ P ∥ + ε / 2 < ε ( 1 ≤ j ≤ s ) \lVert\theta_j|_{D'\cap P}\rVert<\lVert\theta_j|_{F'\cap P}\rVert+\varepsilon/2<\varepsilon\qquad(1\le j\le s) ∥ θ j ∣ D ′ ∩ P ∥ < ∥ θ j ∣ F ′ ∩ P ∥ + ε /2 < ε ( 1 ≤ j ≤ s )
(Lemma 6.4 and Proposition 6.5 ).
We use our Poisson boundary theory to construct M M M -inner ucp maps R n R_n R n with finitely many coefficients such that ∥ θ ∘ R n ∥ → 0 \lVert\theta\circ R_n\rVert\to0 ∥ θ ∘ R n ∥ → 0 for every θ ∈ Z ( M ) ⊥ \theta\in Z(M)_\perp θ ∈ Z ( M ) ⊥ . We will use these maps to construct a finite-dimensional matrix algebra F ⊂ M F\subset M F ⊂ M with ∥ θ ∣ F ′ ∩ P ∥ < ε \lVert\theta|_{F'\cap P}\rVert<\varepsilon ∥ θ ∣ F ′ ∩ P ∥ < ε .
Lemma 6.2. Let M ⊂ P M\subset P M ⊂ P be as in Theorem G . Then there is a sequence of normal M M M -inner ucp maps R n : P → P R_n:P\to P R n : P → P , a ↦ ∑ k = 1 r ( n ) c n , k a c n , k ∗ a\mapsto\sum_{k=1}^{r(n)}c_{n,k}ac_{n,k}^* a ↦ ∑ k = 1 r ( n ) c n , k a c n , k ∗ , each with finitely many coefficients c n , k ∈ M c_{n,k}\in M c n , k ∈ M , such that
∥ θ ∘ R n ∥ ⟶ 0 ( n → ∞ ) \lVert\theta\circ R_n\rVert\longrightarrow0 \qquad(n\to\infty) ∥ θ ∘ R n ∥ ⟶ 0 ( n → ∞ ) for every θ ∈ Z ( M ) ⊥ = { θ ∈ P ∗ ∣ θ ∣ Z ( M ) = 0 } \theta\in Z(M)_\perp=\{\theta\in P_* \mid\theta|_{Z(M)}=0\} θ ∈ Z ( M ) ⊥ = { θ ∈ P ∗ ∣ θ ∣ Z ( M ) = 0 } .
Proof. Fix a normal faithful state φ \varphi φ on M M M . Take an nrsg ( M , φ ) (M,\varphi) ( M , φ ) -hyperstate φ \varphi φ , which always exists by Example 3.9 , with inner ucp coefficients ( y k ) ⊂ M (y_k)\subset M ( y k ) ⊂ M . Define the normal M M M -inner ucp map
Q ( x ) = ∑ k y k x y k ∗ ( x ∈ P ) . (55) Q(x)=\sum_k y_kxy_k^* \qquad(x\in P). \tag*{(55)} Q ( x ) = k ∑ y k x y k ∗ ( x ∈ P ) . (55)
By Proposition 3.13, Har ( Q , P ) = M ′ ∩ P = Z ( M ) \operatorname{Har}(Q,P)=M'\cap P=\mathcal{Z}(M) Har ( Q , P ) = M ′ ∩ P = Z ( M ) .
Since P ∗ P_* P ∗ is separable, take a norm-dense sequence ( θ j ) j ≥ 1 (\theta_j)_{j\geq1} ( θ j ) j ≥ 1 in Z ( M ) ⊥ \mathcal{Z}(M)_\perp Z ( M ) ⊥ . Let C n = ( 1 / n ) ∑ ℓ = 0 n − 1 Q ℓ C_n=(1/n)\sum_{\ell=0}^{n-1}Q^\ell C n = ( 1/ n ) ∑ ℓ = 0 n − 1 Q ℓ . Let T = Q ∗ T=Q_* T = Q ∗ be the predual map on P ∗ P_* P ∗ , so that T θ = θ ∘ Q T\theta=\theta\circ Q T θ = θ ∘ Q . The annihilator of Ran ( id − T ) \operatorname{Ran}(\operatorname{id}-T) Ran ( id − T ) in P = ( P ∗ ) ∗ P=(P_*)^* P = ( P ∗ ) ∗ is Har ( Q , P ) = Z ( M ) \operatorname{Har}(Q,P)=\mathcal{Z}(M) Har ( Q , P ) = Z ( M ) . Hence by the Hahn–Banach theorem, we have
Ran ( id − T ) ‾ ∥ ⋅ ∥ = { θ ∈ P ∗ ∣ θ ∣ Z ( M ) = 0 } . \overline{\operatorname{Ran}(\operatorname{id}-T)}^{\|\cdot\|}=\{\theta\in P_* \mid\theta|_{\mathcal{Z}(M)}=0\}. Ran ( id − T ) ∥ ⋅ ∥ = { θ ∈ P ∗ ∣ θ ∣ Z ( M ) = 0 } .
For θ = ( id − T ) η \theta=(\operatorname{id}-T)\eta θ = ( id − T ) η with η ∈ P ∗ \eta\in P_* η ∈ P ∗ , we have
∥ θ ∘ C n ∥ = 1 n ∥ η − T n η ∥ ≤ 2 ∥ η ∥ n . \|\theta\circ C_n\|=\frac{1}{n}\|\eta-T^n\eta\|\leq\frac{2\|\eta\|}{n}. ∥ θ ∘ C n ∥ = n 1 ∥ η − T n η ∥ ≤ n 2∥ η ∥ .
Since ∥ C n ∥ = 1 \|C_n\|=1 ∥ C n ∥ = 1 , this convergence extends to the norm closure of Ran ( id − T ) \operatorname{Ran}(\operatorname{id}-T) Ran ( id − T ) . In particular, ∥ θ j ∘ C n ∥ → 0 \|\theta_j\circ C_n\|\to0 ∥ θ j ∘ C n ∥ → 0 as n → ∞ n\to\infty n → ∞ for every j ≥ 1 j\geq1 j ≥ 1 .
Each C n C_n C n is a normal M M M -inner ucp map. Write C n ( x ) = ∑ i ≥ 1 a n , i x a n , i ∗ C_n(x)=\sum_{i\geq1}a_{n,i}xa_{n,i}^* C n ( x ) = ∑ i ≥ 1 a n , i x a n , i ∗ , ( x ∈ P ) (x\in P) ( x ∈ P ) with a n , i ∈ M a_{n,i}\in M a n , i ∈ M and ∑ i a n , i a n , i ∗ = 1 \sum_i a_{n,i}a_{n,i}^*=1 ∑ i a n , i a n , i ∗ = 1 . For m ≥ 1 m\geq1 m ≥ 1 , define
b n , m = ( 1 − ∑ i = 1 m a n , i a n , i ∗ ) 1 / 2 , R n , m ( x ) = ∑ i = 1 m a n , i x a n , i ∗ + b n , m x b n , m . b_{n,m}=\left(1-\sum_{i=1}^{m}a_{n,i}a_{n,i}^*\right)^{1/2},\qquad R_{n,m}(x)=\sum_{i=1}^{m}a_{n,i}xa_{n,i}^*+b_{n,m}xb_{n,m}. b n , m = ( 1 − i = 1 ∑ m a n , i a n , i ∗ ) 1/2 , R n , m ( x ) = i = 1 ∑ m a n , i x a n , i ∗ + b n , m x b n , m .
Then R n , m R_{n,m} R n , m is an M M M -inner ucp map with finitely many coefficients. For θ ∈ P ∗ + \theta\in P_*^+ θ ∈ P ∗ + , the functional θ ∘ R n , m − θ ∘ C n \theta\circ R_{n,m}-\theta\circ C_n θ ∘ R n , m − θ ∘ C n is the difference of the positive functionals θ ( b n , m ⋅ b n , m ) \theta(b_{n,m}\,\cdot\,b_{n,m}) θ ( b n , m ⋅ b n , m ) and ∑ i > m θ ( a n , i ⋅ a n , i ∗ ) \sum_{i>m}\theta(a_{n,i}\,\cdot\,a_{n,i}^*) ∑ i > m θ ( a n , i ⋅ a n , i ∗ ) , both of norm θ ( b n , m 2 ) \theta(b_{n,m}^2) θ ( b n , m 2 ) . Hence, for fixed n n n , we have
∥ θ ∘ R n , m − θ ∘ C n ∥ ≤ 2 θ ( b n , m 2 ) ⟶ 0 ( m → ∞ ) . \|\theta\circ R_{n,m}-\theta\circ C_n\|\leq2\theta(b_{n,m}^2)\longrightarrow0\qquad(m\to\infty). ∥ θ ∘ R n , m − θ ∘ C n ∥ ≤ 2 θ ( b n , m 2 ) ⟶ 0 ( m → ∞ ) .
Every normal functional is a linear combination of positive normal functionals, so the convergence holds for every θ ∈ P ∗ \theta\in P_* θ ∈ P ∗ . For each n ≥ 1 n\geq1 n ≥ 1 , take m ( n ) m(n) m ( n ) sufficiently large and let R n = R n , m ( n ) R_n=R_{n,m(n)} R n = R n , m ( n ) so that
∥ θ j ∘ R n − θ j ∘ C n ∥ < 2 − n ( 1 ≤ j ≤ n ) . \|\theta_j\circ R_n-\theta_j\circ C_n\|<2^{-n}\qquad(1\leq j\leq n). ∥ θ j ∘ R n − θ j ∘ C n ∥ < 2 − n ( 1 ≤ j ≤ n ) .
For every fixed j j j and n ≥ j n\geq j n ≥ j , we have
∥ θ j ∘ R n ∥ ≤ 2 − n + ∥ θ j ∘ C n ∥ ⟶ 0 ( n → ∞ ) . \|\theta_j\circ R_n\|\leq2^{-n}+\|\theta_j\circ C_n\|\longrightarrow0\qquad(n\to\infty). ∥ θ j ∘ R n ∥ ≤ 2 − n + ∥ θ j ∘ C n ∥ ⟶ 0 ( n → ∞ ) .
Since ∥ R n ∥ = 1 \|R_n\|=1 ∥ R n ∥ = 1 , norm density gives the same convergence for every θ ∈ Z ( M ) ⊥ \theta\in\mathcal{Z}(M)_\perp θ ∈ Z ( M ) ⊥ .
We use the proper infiniteness of M M M and the maps R n R_n R n constructed above to construct a full matrix algebra F ⊂ M F\subset M F ⊂ M with ∥ θ ∣ F ′ ∩ P ∥ < ε \|\theta|_{F'\cap P}\|<\varepsilon ∥ θ ∣ F ′ ∩ P ∥ < ε .
Proposition 6.3. Let M ⊂ P M\subset P M ⊂ P be as in Theorem G . Assume that M M M is properly infinite. Let θ 1 , … , θ s ∈ Z ( M ) ⊥ \theta_1,\ldots,\theta_s\in\mathcal{Z}(M)_\perp θ 1 , … , θ s ∈ Z ( M ) ⊥ , and let ε > 0 \varepsilon>0 ε > 0 . Then there is a full matrix algebra F ⊂ M F\subset M F ⊂ M such that
∥ θ j ∣ F ′ ∩ P ∥ < ε ( 1 ≤ j ≤ s ) . \|\theta_j|_{F'\cap P}\|<\varepsilon\qquad(1\leq j\leq s). ∥ θ j ∣ F ′ ∩ P ∥ < ε ( 1 ≤ j ≤ s ) .
Proof. By Lemma 6.2 , take a normal M M M -inner ucp map R ( x ) = ∑ i = 1 r a i x a i ∗ R(x)=\sum_{i=1}^{r}a_ix a_i^* R ( x ) = ∑ i = 1 r a i x a i ∗ ( x ∈ P ) (x\in P) ( x ∈ P ) with finitely many coefficients a i ∈ M a_i\in M a i ∈ M such that ∥ θ j ∘ R ∥ < ε / 2 \|\theta_j\circ R\|<\varepsilon/2 ∥ θ j ∘ R ∥ < ε /2 for 1 ≤ j ≤ s 1\leq j\leq s 1 ≤ j ≤ s . We use the same matrix construction as in the proof of Theorem 4.1. Let d = r + 1 d=r+1 d = r + 1 , and fix a projection e ∈ M e\in M e ∈ M with e ∼ 1 − e ∼ 1 e\sim1-e\sim1 e ∼ 1 − e ∼ 1 . Take matrix units ( e i j ) 1 ≤ i , j ≤ d (e_{ij})_{1\leq i,j\leq d} ( e ij ) 1 ≤ i , j ≤ d in M d ( C ) ⊗ 1 M_d(\mathbb{C})\otimes1 M d ( C ) ⊗ 1 . Take an isometry v e ∈ M v_e\in M v e ∈ M with v e v e ∗ = e v_ev_e^*=e v e v e ∗ = e and an isomorphism
π e : M ⟶ ∼ M d ( C ) ⊗ ‾ M , π e ( e ) = e 11 , \pi_e:M\overset{\sim}{\longrightarrow}M_d(\mathbb{C})\overline{\otimes}M,\qquad\pi_e(e)=e_{11}, π e : M ⟶ ∼ M d ( C ) ⊗ M , π e ( e ) = e 11 ,
which extends to an isomorphism P ⟶ ∼ M d ( C ) ⊗ ‾ P P \overset{\sim}{\longrightarrow} M_d(\mathbb{C})\overline{\otimes}P P ⟶ ∼ M d ( C ) ⊗ P , still denoted by π e \pi_e π e , such that
π e − 1 ( ( x 0 1 × r 0 r × 1 0 r × r ) ) = v e x v e ∗ ( x ∈ P ) . \pi_e^{-1}\left(\begin{pmatrix} x & 0_{1\times r} \\ 0_{r\times1} & 0_{r\times r} \end{pmatrix}\right)=v_e x v_e^*
\qquad(x\in P). π e − 1 ( ( x 0 r × 1 0 1 × r 0 r × r ) ) = v e x v e ∗ ( x ∈ P ) .
In M d ( C ) ⊗ ‾ M M_d(\mathbb{C})\overline{\otimes}M M d ( C ) ⊗ M , let
Z = ( 0 a 1 ⋯ a r 0 0 ⋯ 0 ⋮ ⋮ ⋱ ⋮ 0 0 ⋯ 0 ) . Z=\begin{pmatrix}
0 & a_1 & \cdots& a_r\\
0 & 0 & \cdots& 0\\
\vdots& \vdots& \ddots& \vdots\\
0 & 0 & \cdots& 0
\end{pmatrix}. Z = 0 0 ⋮ 0 a 1 0 ⋮ 0 ⋯ ⋯ ⋱ ⋯ a r 0 ⋮ 0 .
Since ∑ i = 1 r a i a i ∗ = 1 \sum_{i=1}^{r}a_i a_i^*=1 ∑ i = 1 r a i a i ∗ = 1 and the first column of Z Z Z is 0 0 0 ,
Z Z ∗ = e 11 , 1 − Z ∗ Z ≥ e 11 ∼ 1 , 1 − e 11 ∼ 1. ZZ^*=e_{11},\qquad1-Z^*Z\geq e_{11}\sim1,\qquad1-e_{11}\sim1. Z Z ∗ = e 11 , 1 − Z ∗ Z ≥ e 11 ∼ 1 , 1 − e 11 ∼ 1.
Take a partial isometry W ∈ M d ( C ) ⊗ ‾ M W\in M_d(\mathbb{C})\overline{\otimes}M W ∈ M d ( C ) ⊗ M with
W ∗ W = 1 − Z ∗ Z , W W ∗ = 1 − e 11 . W^*W=1-Z^*Z,\qquad WW^*=1-e_{11}. W ∗ W = 1 − Z ∗ Z , W W ∗ = 1 − e 11 .
Then U = Z + W U=Z+W U = Z + W is unitary and its first row is ( 0 , a 1 , … , a r ) (0,a_1,\ldots,a_r) ( 0 , a 1 , … , a r ) . For x ∈ P x\in P x ∈ P , by matrix multiplication, we have
U ( x 0 ⋯ 0 0 x ⋯ 0 ⋮ ⋮ ⋱ ⋮ 0 0 ⋯ x ) U ∗ = ( R ( x ) ∗ ∗ ∗ ) . U\begin{pmatrix}
x & 0 & \cdots& 0\\
0 & x & \cdots& 0\\
\vdots& \vdots& \ddots& \vdots\\
0 & 0 & \cdots& x
\end{pmatrix}U^*
=
\begin{pmatrix}
R(x) & *\\
* & *
\end{pmatrix}. U x 0 ⋮ 0 0 x ⋮ 0 ⋯ ⋯ ⋱ ⋯ 0 0 ⋮ x U ∗ = ( R ( x ) ∗ ∗ ∗ ) .
Define
α e ( x ) = π e − 1 ( U ( 1 d ⊗ x ) U ∗ ) , F e = π e − 1 ( U ( M d ( C ) ⊗ 1 ) U ∗ ) . \alpha_e(x)=\pi_e^{-1}\left(U(1_d\otimes x)U^*\right),\qquad
F_e=\pi_e^{-1}\left(U(M_d(\mathbb{C})\otimes1)U^*\right). α e ( x ) = π e − 1 ( U ( 1 d ⊗ x ) U ∗ ) , F e = π e − 1 ( U ( M d ( C ) ⊗ 1 ) U ∗ ) .
Then α e \alpha_e α e is a normal unital injective endomorphism, F e ⊂ M F_e\subset M F e ⊂ M is a full matrix algebra, and
α e ( P ) = F e ′ ∩ P . \alpha_e(P)=F_e'\cap P. α e ( P ) = F e ′ ∩ P .
Since π e ( e ) = e 11 \pi_e(e)=e_{11} π e ( e ) = e 11 , we have
e α e ( x ) e = π e − 1 ( ( R ( x ) 0 1 × r 0 r × 1 0 r × r ) ) = v e R ( x ) v e ∗ ( x ∈ P ) . (56) e\alpha_e(x)e=\pi_e^{-1}\left(\begin{pmatrix} R(x) & 0_{1\times r}\\ 0_{r\times1} & 0_{r\times r}\end{pmatrix}\right)=v_eR(x)v_e^*
\qquad(x\in P). \tag*{(56)} e α e ( x ) e = π e − 1 ( ( R ( x ) 0 r × 1 0 1 × r 0 r × r ) ) = v e R ( x ) v e ∗ ( x ∈ P ) . (56)
Finally, take projections e n ↗ 1 e_n\nearrow1 e n ↗ 1 strongly in M M M , with e 0 = 0 e_0=0 e 0 = 0 and e n − e n − 1 ∼ 1 − e n ∼ 1 e_n-e_{n-1}\sim1-e_n\sim1 e n − e n − 1 ∼ 1 − e n ∼ 1 for n ≥ 1 n\geq1 n ≥ 1 . Take partial isometries v n ∈ M v_n\in M v n ∈ M with v n ∗ v n = 1 − e n − 1 v_n^*v_n=1-e_{n-1} v n ∗ v n = 1 − e n − 1 and v n v n ∗ = e n − e n − 1 v_nv_n^*=e_n-e_{n-1} v n v n ∗ = e n − e n − 1 , and use v e n = e n − 1 + v n v_{e_n}=e_{n-1}+v_n v e n = e n − 1 + v n in the construction above. Since v e n e n − 1 = v e n − 1 = e n − 1 v_{e_n}e_{n-1}=v_{e_{n-1}}=e_{n-1} v e n e n − 1 = v e n − 1 = e n − 1 , we have v e n → 1 v_{e_n}\to1 v e n → 1 *-strongly. Let α n = α e n \alpha_n=\alpha_{e_n} α n = α e n and F n = F e n F_n=F_{e_n} F n = F e n . For θ ∈ P ∗ \theta\in P_* θ ∈ P ∗ , by (56) , we have
∥ θ ∘ α n − θ ∘ R ∥ ≤ ∥ θ − θ ( e n ⋅ e n ) ∥ + ∥ θ ( v e n ⋅ v e n ∗ ) − θ ∥ ⟶ 0 ( n → ∞ ) . \|\theta\circ\alpha_n-\theta\circ R\|\leq\|\theta-\theta(e_n\cdot e_n)\|+\|\theta(v_{e_n}\cdot v_{e_n}^*)-\theta\|\longrightarrow0
\qquad(n\to\infty). ∥ θ ∘ α n − θ ∘ R ∥ ≤ ∥ θ − θ ( e n ⋅ e n ) ∥ + ∥ θ ( v e n ⋅ v e n ∗ ) − θ ∥ ⟶ 0 ( n → ∞ ) .
The convergence follows first for normal states by the Cauchy–Schwarz inequality and then for every θ ∈ P ∗ \theta\in P_* θ ∈ P ∗ by linearity. Take n n n sufficiently large that ∥ θ j ∘ α n − θ j ∘ R ∥ < ε / 2 \|\theta_j\circ\alpha_n-\theta_j\circ R\|<\varepsilon/2 ∥ θ j ∘ α n − θ j ∘ R ∥ < ε /2 for 1 ≤ j ≤ s 1\leq j\leq s 1 ≤ j ≤ s . Since α n \alpha_n α n is a *-isomorphism of P P P onto F n ′ ∩ P F_n'\cap P F n ′ ∩ P , we have
∥ θ j ∣ F n ′ ∩ P ∥ = ∥ θ j ∘ α n ∥ ≤ ∥ θ j ∘ α n − θ j ∘ R ∥ + ∥ θ j ∘ R ∥ < ε ( 1 ≤ j ≤ s ) . \|\theta_j|_{F_n'\cap P}\|=\|\theta_j\circ\alpha_n\|\leq\|\theta_j\circ\alpha_n-\theta_j\circ R\|+\|\theta_j\circ R\|<\varepsilon
\qquad(1\leq j\leq s). ∥ θ j ∣ F n ′ ∩ P ∥ = ∥ θ j ∘ α n ∥ ≤ ∥ θ j ∘ α n − θ j ∘ R ∥ + ∥ θ j ∘ R ∥ < ε ( 1 ≤ j ≤ s ) .
Taking F = F n F=F_n F = F n finishes the proof. □
For a finite-dimensional abelian subalgebra D ⊂ P D\subset P D ⊂ P , let At ( D ) \operatorname{At}(D) At ( D ) denote the set of minimal projections of D D D , and define
Φ D ( x ) = ∑ p ∈ At ( D ) p x p ( x ∈ P ) . \Phi_D(x)=\sum_{p\in\operatorname{At}(D)}pxp
\qquad(x\in P). Φ D ( x ) = p ∈ At ( D ) ∑ p x p ( x ∈ P ) .
For a finite-dimensional subalgebra F ⊂ P F \subset P F ⊂ P , define
Ψ F ( x ) = ∫ U ( F ) u x u ∗ d u ( x ∈ P ) , \Psi_F(x)=\int_{U(F)}uxu^*\,du \qquad(x\in P), Ψ F ( x ) = ∫ U ( F ) ux u ∗ d u ( x ∈ P ) ,
where d u du d u is the normalized Haar measure on U ( F ) U(F) U ( F ) . Then Φ D \Phi_D Φ D and Ψ F \Psi_F Ψ F are the normal conditional expectations onto D ′ ∩ P D'\cap P D ′ ∩ P and F ′ ∩ P F'\cap P F ′ ∩ P , respectively. In particular,
∥ θ ∣ D ′ ∩ P ∥ = ∥ θ ∘ Φ D ∥ , ∥ θ ∣ F ′ ∩ P ∥ = ∥ θ ∘ Ψ F ∥ ( θ ∈ P ∗ ) . \|\theta|_{D'\cap P}\|=\|\theta\circ\Phi_D\|,\qquad\|\theta|_{F'\cap P}\|=\|\theta\circ\Psi_F\| \qquad(\theta\in P_*). ∥ θ ∣ D ′ ∩ P ∥ = ∥ θ ∘ Φ D ∥ , ∥ θ ∣ F ′ ∩ P ∥ = ∥ θ ∘ Ψ F ∥ ( θ ∈ P ∗ ) .
The following simple lemma will help us pass from the matrix algebra F F F to an abelian subalgebra D D D with ∥ θ ∣ D ′ ∩ P ∥ < ε \|\theta|_{D'\cap P}\|<\varepsilon ∥ θ ∣ D ′ ∩ P ∥ < ε .
Lemma 6.4. Let M ⊂ P M\subset P M ⊂ P be as in Theorem G . Let F ⊂ P F\subset P F ⊂ P be a full matrix algebra with F ≅ M d ( C ) F\cong M_d(\mathbb{C}) F ≅ M d ( C ) , and let D ⊂ F D\subset F D ⊂ F be a maximal abelian subalgebra. Then
∫ U ( F ) Φ u D u ∗ d u = id + d Ψ F d + 1 . (57) \int_{U(F)}\Phi_{uDu^*}\,du=\frac{\operatorname{id}+d\Psi_F}{d+1}. \tag*{(57)} ∫ U ( F ) Φ u D u ∗ d u = d + 1 id + d Ψ F . (57) If ρ \rho ρ is a normal faithful state on P P P with F ⊂ P ρ F\subset P_\rho F ⊂ P ρ , then
∫ U ( F ) ∥ Φ u D u ∗ ( b ) ∥ 2 , ρ 2 d u = ∥ b ∥ 2 , ρ 2 d + 1 \int_{U(F)}\|\Phi_{uDu^*}(b)\|_{2,\rho}^2\,du=\frac{\|b\|_{2,\rho}^2}{d+1} ∫ U ( F ) ∥ Φ u D u ∗ ( b ) ∥ 2 , ρ 2 d u = d + 1 ∥ b ∥ 2 , ρ 2 for every b ∈ P b\in P b ∈ P with Ψ F ( b ) = 0 \Psi_F(b)=0 Ψ F ( b ) = 0 .
Proof. Under the identification P = F ⊗ ‾ ( F ′ ∩ P ) P=F\mathbin{\overline{\otimes}}(F'\cap P) P = F ⊗ ( F ′ ∩ P ) , both sides of (57) act only on the first tensor factor. Hence we only need to prove the identity on F ≅ M d ( C ) F\cong M_d(\mathbb{C}) F ≅ M d ( C ) . The case d = 1 d=1 d = 1 is immediate. Let Φ = ∫ U ( F ) Φ u D u ∗ d u \Phi=\int_{U(F)}\Phi_{uDu^*}\,du Φ = ∫ U ( F ) Φ u D u ∗ d u . Let p ∈ F p\in F p ∈ F be a rank-one projection. For any u ∈ U ( { p } ′ ∩ F ) u\in U(\{p\}'\cap F) u ∈ U ({ p } ′ ∩ F ) , invariance of Haar measure gives
u Φ ( p ) u ∗ = Φ ( u p u ∗ ) = Φ ( p ) . u\Phi(p)u^*=\Phi(upu^*)=\Phi(p). u Φ ( p ) u ∗ = Φ ( u p u ∗ ) = Φ ( p ) .
Hence Φ ( p ) ∈ ( { p } ′ ∩ F ) ′ ∩ F = { p } ′ ′ ∩ F = C p + C ( 1 − p ) \Phi(p)\in(\{p\}'\cap F)'\cap F=\{p\}''\cap F=\mathbb{C}p+\mathbb{C}(1-p) Φ ( p ) ∈ ({ p } ′ ∩ F ) ′ ∩ F = { p } ′′ ∩ F = C p + C ( 1 − p ) . Since all rank-one projections are unitarily conjugate and span F F F , we have Φ = λ id + μ tr d ( ⋅ ) 1 \Phi=\lambda\operatorname{id}+\mu\operatorname{tr}_d(\cdot)1 Φ = λ id + μ tr d ( ⋅ ) 1 for some scalars λ , μ \lambda,\mu λ , μ , where tr d \operatorname{tr}_d tr d is the normalized trace on F F F . Each Φ u D u ∗ \Phi_{uDu^*} Φ u D u ∗ is a rank-d d d orthogonal projection in B ( L 2 ( F , tr d ) ) B(L^2(F,\operatorname{tr}_d)) B ( L 2 ( F , tr d )) . Taking the operator trace Tr \operatorname{Tr} Tr on B ( L 2 ( F , tr d ) ) B(L^2(F,\operatorname{tr}_d)) B ( L 2 ( F , tr d )) and using Φ ( 1 ) = 1 \Phi(1)=1 Φ ( 1 ) = 1 , we have
d 2 λ + μ = d , λ + μ = 1. d^2\lambda+\mu=d,\qquad\lambda+\mu=1. d 2 λ + μ = d , λ + μ = 1.
Hence λ = 1 / ( d + 1 ) \lambda=1/(d+1) λ = 1/ ( d + 1 ) and μ = d / ( d + 1 ) \mu=d/(d+1) μ = d / ( d + 1 ) . Tensoring with the identity on F ′ ∩ P F'\cap P F ′ ∩ P proves (57) .
Assume now that F ⊂ P ρ F\subset P_\rho F ⊂ P ρ . Then each Φ u D u ∗ \Phi_{uDu^*} Φ u D u ∗ is the ρ \rho ρ -preserving conditional expectation onto ( u D u ∗ ) ′ ∩ P (uDu^*)'\cap P ( u D u ∗ ) ′ ∩ P , and hence acts as an orthogonal projection on L 2 ( P , ρ ) L^2(P,\rho) L 2 ( P , ρ ) . For b ∈ P b\in P b ∈ P with Ψ F ( b ) = 0 \Psi_F(b)=0 Ψ F ( b ) = 0 , by (57) we have
∫ U ( F ) ∥ Φ u D u ∗ ( b ) ∥ 2 , ρ 2 d u = ∫ U ( F ) ⟨ Φ u D u ∗ ( b ) , b ⟩ 2 , ρ d u = ⟨ b + d Ψ F ( b ) , b ⟩ 2 , ρ d + 1 = ∥ b ∥ 2 , ρ 2 d + 1 . \int_{U(F)}\|\Phi_{uDu^*}(b)\|_{2,\rho}^2\,du
=\int_{U(F)}\langle\Phi_{uDu^*}(b),b\rangle_{2,\rho}\,du
=\frac{\langle b+d\Psi_F(b),b\rangle_{2,\rho}}{d+1}
=\frac{\|b\|_{2,\rho}^2}{d+1}. ∫ U ( F ) ∥ Φ u D u ∗ ( b ) ∥ 2 , ρ 2 d u = ∫ U ( F ) ⟨ Φ u D u ∗ ( b ) , b ⟩ 2 , ρ d u = d + 1 ⟨ b + d Ψ F ( b ) , b ⟩ 2 , ρ = d + 1 ∥ b ∥ 2 , ρ 2 .
This proves (6.5). □
We now construct the required finite-dimensional abelian subalgebra when M M M is continuous and properly infinite.
Proposition 6.5. Let M ⊂ P M\subset P M ⊂ P be as in Theorem G . Assume that M M M is continuous and properly infinite. Let θ 1 , … , θ s ∈ M ⊥ \theta_1,\ldots,\theta_s\in M_\perp θ 1 , … , θ s ∈ M ⊥ , and let ε > 0 \varepsilon>0 ε > 0 . Then there is a finite-dimensional abelian subalgebra D ⊂ M D\subset M D ⊂ M such that
∥ θ j ∣ D ′ ∩ P ∥ = ∥ θ j ∘ Φ D ∥ < ε ( 1 ≤ j ≤ s ) . \|\theta_j|_{D'\cap P}\|=\|\theta_j\circ\Phi_D\|<\varepsilon\qquad(1\leq j\leq s). ∥ θ j ∣ D ′ ∩ P ∥ = ∥ θ j ∘ Φ D ∥ < ε ( 1 ≤ j ≤ s ) .
Proof. By Proposition 6.3, take a full matrix algebra F ⊂ M F \subset M F ⊂ M such that ∥ θ j ∣ F ′ ∩ P ∥ < ε / 2 \|\theta_j|_{F'\cap P}\| < \varepsilon/2 ∥ θ j ∣ F ′ ∩ P ∥ < ε /2 for 1 ≤ j ≤ s 1 \le j \le s 1 ≤ j ≤ s . Let F ≅ M r ( C ) F \cong M_r(\mathbb{C}) F ≅ M r ( C ) and N = F ′ ∩ M N=F'\cap M N = F ′ ∩ M . Then M = F ⊗ ‾ N M=F\overline{\otimes}N M = F ⊗ N , and N N N is a continuous von Neumann algebra. By [22 Theorem 11.1] , take a normal faithful state χ \chi χ on N N N such that N χ N_\chi N χ is of type I I 1 \mathrm{II}_1 II 1 and hence contains a copy of M k ( C ) M_k(\mathbb{C}) M k ( C ) for every k ≥ 1 k\ge1 k ≥ 1 . Indeed, there exists a projection p ∈ N χ p\in N_\chi p ∈ N χ with χ ( p ) = 1 / k \chi(p)=1/k χ ( p ) = 1/ k such that N χ ≅ M k ( C ) ⊗ ‾ p N χ p N_\chi\cong M_k(\mathbb{C})\overline{\otimes}pN_\chi p N χ ≅ M k ( C ) ⊗ p N χ p . Let tr r \operatorname{tr}_r tr r be the normalized trace on M r ( C ) M_r(\mathbb{C}) M r ( C ) , and let
ψ = tr r ⊗ ‾ χ ∈ M ∗ , ρ = ψ ∘ E ∈ P ∗ . \psi=\operatorname{tr}_r\overline{\otimes}\chi\in M_*,\qquad\rho=\psi\circ E\in P_*. ψ = tr r ⊗ χ ∈ M ∗ , ρ = ψ ∘ E ∈ P ∗ .
Since ρ \rho ρ is faithful, the only x ∈ P x\in P x ∈ P with ρ ( a x ) = 0 \rho(ax)=0 ρ ( a x ) = 0 for all a ∈ P a\in P a ∈ P is x = 0 x=0 x = 0 . By the Hahn–Banach theorem,
{ ρ ( a ⋅ ) ∣ a ∈ P } ‾ ∥ ⋅ ∥ = P ∗ . (58) \overline{\{\rho(a\cdot)\mid a\in P\}}^{\|\cdot\|}=P_*. \tag*{(58)} { ρ ( a ⋅ ) ∣ a ∈ P } ∥ ⋅ ∥ = P ∗ . (58)
By (58) , we can take a 1 , … , a s ∈ P a_1,\ldots,a_s\in P a 1 , … , a s ∈ P and k ≥ 1 k\ge1 k ≥ 1 such that
∥ θ j − ρ ( a j ⋅ ) ∥ < ε 8 , ( 4 ∑ j = 1 s ∥ a j ∥ 2 r k + 1 ) 1 / 2 < ε 4 . \|\theta_j-\rho(a_j\cdot)\|<\frac{\varepsilon}{8},\qquad\left(\frac{4\sum_{j=1}^{s}\|a_j\|^2}{rk+1}\right)^{1/2}<\frac{\varepsilon}{4}. ∥ θ j − ρ ( a j ⋅ ) ∥ < 8 ε , ( r k + 1 4 ∑ j = 1 s ∥ a j ∥ 2 ) 1/2 < 4 ε .
Let K ⊂ N χ K\subset N_\chi K ⊂ N χ be a subalgebra with K ≅ M k ( C ) K\cong M_k(\mathbb{C}) K ≅ M k ( C ) , and let F 1 = F ∨ K F_1=F\vee K F 1 = F ∨ K . Then
F 1 ≅ M r k ( C ) , F 1 ⊂ M ψ ⊂ P ρ . F_1\cong M_{rk}(\mathbb{C}),\qquad F_1\subset M_\psi\subset P_\rho. F 1 ≅ M r k ( C ) , F 1 ⊂ M ψ ⊂ P ρ .
Let c j = a j − Ψ F ( a j ) c_j=a_j-\Psi_F(a_j) c j = a j − Ψ F ( a j ) for 1 ≤ j ≤ s 1\le j\le s 1 ≤ j ≤ s . Since U ( F ) ⊂ U ( F 1 ) U(F)\subset U(F_1) U ( F ) ⊂ U ( F 1 ) , by the invariance of Haar measure we have Ψ F 1 Ψ F = Ψ F 1 \Psi_{F_1}\Psi_F=\Psi_{F_1} Ψ F 1 Ψ F = Ψ F 1 , and hence Ψ F 1 ( c j ∗ ) = 0 \Psi_{F_1}(c_j^*)=0 Ψ F 1 ( c j ∗ ) = 0 . Fix a maximal abelian subalgebra A 1 ⊂ F 1 A_1\subset F_1 A 1 ⊂ F 1 . Since ρ ( 1 ) = 1 \rho(1)=1 ρ ( 1 ) = 1 and ∥ Ψ F ∥ = 1 \|\Psi_F\|=1 ∥ Ψ F ∥ = 1 , we have
∥ c j ∗ ∥ 2 , ρ ≤ ∥ c j ∥ ≤ ∥ a j ∥ + ∥ Ψ F ( a j ) ∥ ≤ 2 ∥ a j ∥ . \|c_j^*\|_{2,\rho}\le\|c_j\|\le\|a_j\|+\|\Psi_F(a_j)\|\le2\|a_j\|. ∥ c j ∗ ∥ 2 , ρ ≤ ∥ c j ∥ ≤ ∥ a j ∥ + ∥ Ψ F ( a j ) ∥ ≤ 2∥ a j ∥.
Using F 1 ≅ M r k ( C ) F_1\cong M_{rk}(\mathbb{C}) F 1 ≅ M r k ( C ) , F 1 ⊂ P ρ F_1\subset P_\rho F 1 ⊂ P ρ and Ψ F 1 ( c j ∗ ) = 0 \Psi_{F_1}(c_j^*)=0 Ψ F 1 ( c j ∗ ) = 0 , by (6.5) we have
∫ U ( F 1 ) ∑ j = 1 s ∥ Φ u A 1 u ∗ ( c j ∗ ) ∥ 2 , ρ 2 d u = ∑ j = 1 s ∫ U ( F 1 ) ∥ Φ u A 1 u ∗ ( c j ∗ ) ∥ 2 , ρ 2 d u = ∑ j = 1 s ∥ c j ∗ ∥ 2 , ρ 2 r k + 1 ≤ 4 ∑ j = 1 s ∥ a j ∥ 2 r k + 1 < ε 2 16 . \int_{U(F_1)}\sum_{j=1}^{s}\|\Phi_{uA_1u^*}(c_j^*)\|_{2,\rho}^{2}\,du
=\sum_{j=1}^{s}\int_{U(F_1)}\|\Phi_{uA_1u^*}(c_j^*)\|_{2,\rho}^{2}\,du
=\frac{\sum_{j=1}^{s}\|c_j^*\|_{2,\rho}^{2}}{rk+1}
\le\frac{4\sum_{j=1}^{s}\|a_j\|^{2}}{rk+1}<\frac{\varepsilon^{2}}{16}. ∫ U ( F 1 ) j = 1 ∑ s ∥ Φ u A 1 u ∗ ( c j ∗ ) ∥ 2 , ρ 2 d u = j = 1 ∑ s ∫ U ( F 1 ) ∥ Φ u A 1 u ∗ ( c j ∗ ) ∥ 2 , ρ 2 d u = r k + 1 ∑ j = 1 s ∥ c j ∗ ∥ 2 , ρ 2 ≤ r k + 1 4 ∑ j = 1 s ∥ a j ∥ 2 < 16 ε 2 .
Since d u du d u has total mass 1, there is u ∈ U ( F 1 ) u\in U(F_1) u ∈ U ( F 1 ) for which the sum is no greater than its average. Hence the finite-dimensional abelian subalgebra D = u A 1 u ∗ ⊂ M D=uA_1u^*\subset M D = u A 1 u ∗ ⊂ M satisfies
∑ j = 1 s ∥ Φ D ( c j ∗ ) ∥ 2 , ρ 2 ≤ ∑ j = 1 s ∥ c j ∗ ∥ 2 , ρ 2 r k + 1 < ε 2 16 . \sum_{j=1}^{s}\|\Phi_D(c_j^*)\|_{2,\rho}^{2}\le\frac{\sum_{j=1}^{s}\|c_j^*\|_{2,\rho}^{2}}{rk+1}<\frac{\varepsilon^{2}}{16}. j = 1 ∑ s ∥ Φ D ( c j ∗ ) ∥ 2 , ρ 2 ≤ r k + 1 ∑ j = 1 s ∥ c j ∗ ∥ 2 , ρ 2 < 16 ε 2 .
Since F , D ⊂ P ρ F,D\subset P_\rho F , D ⊂ P ρ , the conditional expectations Ψ F \Psi_F Ψ F and Φ D \Phi_D Φ D preserve ρ \rho ρ , and by bimodularity we have
ρ ( c j ⋅ ) = ρ ( a j ⋅ ) ∘ ( id − Ψ F ) , ρ ( c j ⋅ ) ∘ Φ D = ρ ( Φ D ( c j ) ⋅ ) . \rho(c_j\cdot)=\rho(a_j\cdot)\circ(\operatorname{id}-\Psi_F),\qquad
\rho(c_j\cdot)\circ\Phi_D=\rho(\Phi_D(c_j)\cdot). ρ ( c j ⋅ ) = ρ ( a j ⋅ ) ∘ ( id − Ψ F ) , ρ ( c j ⋅ ) ∘ Φ D = ρ ( Φ D ( c j ) ⋅ ) .
Hence
θ j ∘ Φ D = θ j ∘ Ψ F ∘ Φ D + ρ ( Φ D ( c j ) ⋅ ) + ( θ j − ρ ( a j ⋅ ) ) ∘ ( id − Ψ F ) ∘ Φ D . (59) \theta_j\circ\Phi_D=\theta_j\circ\Psi_F\circ\Phi_D+\rho(\Phi_D(c_j)\cdot)+(\theta_j-\rho(a_j\cdot))\circ(\operatorname{id}-\Psi_F)\circ\Phi_D. \tag*{(59)} θ j ∘ Φ D = θ j ∘ Ψ F ∘ Φ D + ρ ( Φ D ( c j ) ⋅ ) + ( θ j − ρ ( a j ⋅ )) ∘ ( id − Ψ F ) ∘ Φ D . (59)
By (59) , the triangle inequality, and the Cauchy–Schwarz inequality,
∥ θ j ∣ D ′ ∩ P ∥ = ∥ θ j ∘ Φ D ∥ ≤ ∥ θ j ∘ Ψ F ∥ + ∥ Φ D ( c j ∗ ) ∥ 2 , ρ + 2 ∥ θ j − ρ ( a j ⋅ ) ∥ < ∥ θ j ∣ F ′ ∩ P ∥ + ε 2 < ε ( 1 ≤ j ≤ s ) . \|\theta_j|_{D'\cap P}\|=\|\theta_j\circ\Phi_D\|
\le\|\theta_j\circ\Psi_F\|+\|\Phi_D(c_j^*)\|_{2,\rho}+2\|\theta_j-\rho(a_j\cdot)\|
<\|\theta_j|_{F'\cap P}\|+\frac{\varepsilon}{2}<\varepsilon
\qquad(1\le j\le s). ∥ θ j ∣ D ′ ∩ P ∥ = ∥ θ j ∘ Φ D ∥ ≤ ∥ θ j ∘ Ψ F ∥ + ∥ Φ D ( c j ∗ ) ∥ 2 , ρ + 2∥ θ j − ρ ( a j ⋅ ) ∥ < ∥ θ j ∣ F ′ ∩ P ∥ + 2 ε < ε ( 1 ≤ j ≤ s ) .
The finite case
For finite M M M , the proof of the following proposition is essentially Popa’s averaging and refinement argument [38 ] . We include it for completeness, using the tracial case of Example 3.9 to express the unitary averaging through an inner ucp map.
Proposition 6.6. Let M ⊂ P M \subset P M ⊂ P be as in Theorem G . Assume that M M M is finite. Let θ 1 , … , θ s ∈ M ⊥ \theta_1,\ldots,\theta_s \in M_{\perp} θ 1 , … , θ s ∈ M ⊥ , and let ε > 0 \varepsilon> 0 ε > 0 . Then there is a finite-dimensional abelian subalgebra D ⊂ M D \subset M D ⊂ M such that
∥ θ j ∣ D ′ ∩ P ∥ = ∥ θ j ∘ Φ D ∥ < ε ( 1 ≤ j ≤ s ) . \lVert\theta_j|_{D'\cap P}\rVert=\lVert\theta_j\circ\Phi_D\rVert<\varepsilon\qquad(1\leq j\leq s). ∥ θ j ∣ D ′ ∩ P ∥ = ∥ θ j ∘ Φ D ∥ < ε ( 1 ≤ j ≤ s ) .
Proof. Let τ \tau τ be a normal faithful tracial state on M M M , and let ρ = τ ∘ E \rho=\tau\circ E ρ = τ ∘ E . Since M τ = M M_\tau=M M τ = M , we have M ⊂ P ρ M\subset P_\rho M ⊂ P ρ . We first reduce the proposition to an L 2 L^2 L 2 estimate. Since ρ \rho ρ is faithful, we still have (58) . Hence we can take a 1 , … , a s ∈ P a_1,\ldots,a_s\in P a 1 , … , a s ∈ P such that
∥ θ j − ρ ( a j ∗ ⋅ ) ∥ < ε 8 . \lVert\theta_j-\rho(a_j^*\mathbin{\cdot})\rVert<\frac{\varepsilon}{8}. ∥ θ j − ρ ( a j ∗ ⋅ )∥ < 8 ε .
Let b j = a j − E ( a j ) b_j=a_j-E(a_j) b j = a j − E ( a j ) and η j = ρ ( b j ∗ ⋅ ) \eta_j=\rho(b_j^*\mathbin{\cdot}) η j = ρ ( b j ∗ ⋅ ) . Since ρ ∘ E = ρ \rho\circ E=\rho ρ ∘ E = ρ and E E E is M M M -bimodular,
ρ ( a j ∗ E ( x ) ) = ρ ( E ( a j ) ∗ x ) ( x ∈ P ) . \rho(a_j^*E(x))=\rho(E(a_j)^*x)\qquad(x\in P). ρ ( a j ∗ E ( x )) = ρ ( E ( a j ) ∗ x ) ( x ∈ P ) .
Using θ j ∘ E = 0 \theta_j\circ E=0 θ j ∘ E = 0 and ∥ id − E ∥ ≤ 2 \lVert\operatorname{id}-E\rVert\leq2 ∥ id − E ∥ ≤ 2 , we have
E ( b j ) = 0 , θ j − η j = ( θ j − ρ ( a j ∗ ⋅ ) ) ∘ ( id − E ) , ∥ θ j − η j ∥ < ε 4 . E(b_j)=0,\qquad\theta_j-\eta_j=(\theta_j-\rho(a_j^*\mathbin{\cdot}))\circ(\operatorname{id}-E),\qquad\lVert\theta_j-\eta_j\rVert<\frac{\varepsilon}{4}. E ( b j ) = 0 , θ j − η j = ( θ j − ρ ( a j ∗ ⋅ )) ∘ ( id − E ) , ∥ θ j − η j ∥ < 4 ε .
Let D ⊂ M D\subset M D ⊂ M be a finite-dimensional abelian subalgebra. Since D ⊂ M ⊂ P ρ D\subset M\subset P_\rho D ⊂ M ⊂ P ρ , we have
η j ( Φ D ( x ) ) = ∑ p ∈ At ( D ) ρ ( b j ∗ p x p ) = ρ ( Φ D ( b j ) ∗ x ) ( x ∈ P ) . \eta_j(\Phi_D(x))=\sum_{p\in\operatorname{At}(D)}\rho(b_j^*pxp)=\rho(\Phi_D(b_j)^*x)\qquad(x\in P). η j ( Φ D ( x )) = p ∈ At ( D ) ∑ ρ ( b j ∗ p x p ) = ρ ( Φ D ( b j ) ∗ x ) ( x ∈ P ) .
By the Cauchy–Schwarz inequality, we have
∥ ρ ( Φ D ( b j ) ∗ ⋅ ) ∥ ≤ ∥ Φ D ( b j ) ∥ 2 , ρ . \lVert\rho(\Phi_D(b_j)^*\mathbin{\cdot})\rVert\leq\lVert\Phi_D(b_j)\rVert_{2,\rho}. ∥ ρ ( Φ D ( b j ) ∗ ⋅ )∥ ≤ ∥ Φ D ( b j ) ∥ 2 , ρ .
Hence, by the triangle inequality and ∥ Φ D ∥ = 1 \lVert\Phi_D\rVert=1 ∥ Φ D ∥ = 1 ,
∥ θ j ∘ Φ D ∥ ≤ ∥ θ j − η j ∥ + ∥ ρ ( Φ D ( b j ) ∗ ⋅ ) ∥ < ε 4 + ∥ Φ D ( b j ) ∥ 2 , ρ . \lVert\theta_j\circ\Phi_D\rVert\leq\lVert\theta_j-\eta_j\rVert+\lVert\rho(\Phi_D(b_j)^*\mathbin{\cdot})\rVert<\frac{\varepsilon}{4}+\lVert\Phi_D(b_j)\rVert_{2,\rho}. ∥ θ j ∘ Φ D ∥ ≤ ∥ θ j − η j ∥ + ∥ ρ ( Φ D ( b j ) ∗ ⋅ )∥ < 4 ε + ∥ Φ D ( b j ) ∥ 2 , ρ .
Let D \mathcal D D be the family of finite-dimensional abelian subalgebras of M M M , and let
t = inf D ∈ D ∑ j = 1 s ∥ Φ D ( b j ) ∥ 2 , ρ 2 . t=\inf_{D\in\mathcal D}\sum_{j=1}^{s}\lVert\Phi_D(b_j)\rVert_{2,\rho}^{2}. t = D ∈ D inf j = 1 ∑ s ∥ Φ D ( b j ) ∥ 2 , ρ 2 .
We only need to prove that t = 0 t=0 t = 0 , since then some D ∈ D D\in\mathcal D D ∈ D satisfies ∥ Φ D ( b j ) ∥ 2 , ρ < ε / 2 \lVert\Phi_D(b_j)\rVert_{2,\rho}<\varepsilon/2 ∥ Φ D ( b j ) ∥ 2 , ρ < ε /2 for 1 ≤ j ≤ s 1\leq j\leq s 1 ≤ j ≤ s . Fix δ > 0 \delta>0 δ > 0 , and take D ∈ D D\in\mathcal D D ∈ D such that
∑ j = 1 s ∥ Φ D ( b j ) ∥ 2 , ρ 2 < t + δ . \sum_{j=1}^{s}\lVert\Phi_D(b_j)\rVert_{2,\rho}^{2}<t+\delta. j = 1 ∑ s ∥ Φ D ( b j ) ∥ 2 , ρ 2 < t + δ .
Let c j = Φ D ( b j ) c_j=\Phi_D(b_j) c j = Φ D ( b j ) and N = D ′ ∩ M N=D'\cap M N = D ′ ∩ M . By Lemma 6.1(ii), we have N ′ ∩ P = D ∨ Z ( M ) = Z ( N ) ⊂ M N'\cap P=D\vee Z(M)=Z(N)\subset M N ′ ∩ P = D ∨ Z ( M ) = Z ( N ) ⊂ M . Moreover, by the M M M -bimodularity of E E E , we have
E ( c j ) = Φ D ( E ( b j ) ) = 0. E(c_j)=\Phi_D(E(b_j))=0. E ( c j ) = Φ D ( E ( b j )) = 0.
The map Φ D ∘ E : P → N \Phi_D\circ E:P\to N Φ D ∘ E : P → N is a normal faithful conditional expectation. Using Example 3.9 (iii) for ( N , τ ∣ N ) (N,\tau|_N) ( N , τ ∣ N ) , take a normal N N N -inner ucp map Q : P → P Q:P\to P Q : P → P whose coefficients are scalar multiples of unitaries. By the first part of the proof of Lemma 6.2 , its Cesàro means C n = ( 1 / n ) ∑ ℓ = 0 n − 1 Q ℓ C_n=(1/n)\sum_{\ell=0}^{n-1}Q^\ell C n = ( 1/ n ) ∑ ℓ = 0 n − 1 Q ℓ satisfy ∥ θ ∘ C n ∥ → 0 \lVert\theta\circ C_n\rVert\to0 ∥ θ ∘ C n ∥ → 0 as n → ∞ n\to\infty n → ∞ for every θ ∈ N ⊥ = { θ ∈ P ∗ ∣ θ ∣ N = 0 } \theta\in N_{\perp}=\{\theta\in P_* \mid\theta|_N=0\} θ ∈ N ⊥ = { θ ∈ P ∗ ∣ θ ∣ N = 0 } . Each C n C_n C n is a countable convex combination of conjugations by unitaries of N N N . Since E ( c j ) = 0 E(c_j)=0 E ( c j ) = 0 , the functionals ρ ( c j ∗ ⋅ ) \rho(c_j^*\mathbin{\cdot}) ρ ( c j ∗ ⋅ ) vanish on N N N . Hence we have
∑ j = 1 s Re ρ ( c j ∗ C n ( c j ) ) ⟶ 0 ( n → ∞ ) . \sum_{j=1}^{s}\operatorname{Re}\rho(c_j^*C_n(c_j))\longrightarrow0\qquad(n\to\infty). j = 1 ∑ s Re ρ ( c j ∗ C n ( c j )) ⟶ 0 ( n → ∞ ) .
For sufficiently large n n n , one of the unitaries u ∈ U ( N ) u \in U(N) u ∈ U ( N ) in this convex combination satisfies
∑ j = 1 s Re ρ ( c j ∗ u c j u ∗ ) < δ . \sum_{j=1}^{s} \operatorname{Re}\rho(c_j^*uc_ju^*) < \delta. j = 1 ∑ s Re ρ ( c j ∗ u c j u ∗ ) < δ .
Approximating u u u in norm by a finite-spectrum unitary v ∈ W ∗ ( u ) ⊂ N v \in W^*(u) \subset N v ∈ W ∗ ( u ) ⊂ N , we may also arrange
∑ j = 1 s Re ρ ( c j ∗ v c j v ∗ ) < δ . \sum_{j=1}^{s} \operatorname{Re}\rho(c_j^*vc_jv^*) < \delta. j = 1 ∑ s Re ρ ( c j ∗ v c j v ∗ ) < δ .
Let B = D ∨ W ∗ ( v ) B = D \vee W^*(v) B = D ∨ W ∗ ( v ) . Since D D D is finite-dimensional and v ∈ D ′ ∩ M v \in D' \cap M v ∈ D ′ ∩ M has finite spectrum, B B B is a finite-dimensional abelian subalgebra of M M M containing D D D , so B ∈ D B \in\mathcal{D} B ∈ D . Since M ⊂ P ρ M \subset P_\rho M ⊂ P ρ , the map Φ B \Phi_B Φ B is a contraction on L 2 ( P , ρ ) L^2(P,\rho) L 2 ( P , ρ ) and ∥ v c j v ∗ ∥ 2 , ρ = ∥ c j ∥ 2 , ρ \|vc_jv^*\|_{2,\rho} = \|c_j\|_{2,\rho} ∥ v c j v ∗ ∥ 2 , ρ = ∥ c j ∥ 2 , ρ . Using Φ B Φ D = Φ B \Phi_B\Phi_D = \Phi_B Φ B Φ D = Φ B and Φ B ∘ Ad ( v ) = Φ B \Phi_B \circ\operatorname{Ad}(v) = \Phi_B Φ B ∘ Ad ( v ) = Φ B , we have
t ≤ ∑ j ∥ Φ B ( b j ) ∥ 2 , ρ 2 = ∑ j ∥ Φ B ( c j + v c j v ∗ 2 ) ∥ 2 , ρ 2 ≤ 1 2 ∑ j ∥ c j ∥ 2 , ρ 2 + 1 2 ∑ j Re ρ ( c j ∗ v c j v ∗ ) < t 2 + δ . t \leq\sum_j \|\Phi_B(b_j)\|_{2,\rho}^2 = \sum_j \left\|\Phi_B\left(\frac{c_j+vc_jv^*}{2}\right)\right\|_{2,\rho}^2 \leq\frac{1}{2}\sum_j \|c_j\|_{2,\rho}^2 + \frac{1}{2}\sum_j \operatorname{Re}\rho(c_j^*vc_jv^*) < \frac{t}{2}+\delta. t ≤ j ∑ ∥ Φ B ( b j ) ∥ 2 , ρ 2 = j ∑ Φ B ( 2 c j + v c j v ∗ ) 2 , ρ 2 ≤ 2 1 j ∑ ∥ c j ∥ 2 , ρ 2 + 2 1 j ∑ Re ρ ( c j ∗ v c j v ∗ ) < 2 t + δ .
Letting δ → 0 \delta\to0 δ → 0 , we have t = 0 t=0 t = 0 , which proves the proposition. □
Proof of Theorem G
We are ready to put everything together.
Proposition 6.7. Let M ⊂ P M \subset P M ⊂ P be as in Theorem G . Assume that M M M is continuous, i.e., M = M I I ⊕ M I I I M = M_{\mathrm{II}} \oplus M_{\mathrm{III}} M = M II ⊕ M III . Let D 0 ⊂ M D_0 \subset M D 0 ⊂ M be a finite-dimensional abelian subalgebra, let θ 1 , … , θ s ∈ M ⊥ \theta_1,\ldots,\theta_s \in M_\perp θ 1 , … , θ s ∈ M ⊥ , and let ε > 0 \varepsilon> 0 ε > 0 . Then there is a finite-dimensional abelian subalgebra D 1 ⊂ M D_1 \subset M D 1 ⊂ M such that
D 0 ⊂ D 1 and ∥ θ j ∣ D 1 ′ ∩ P ∥ < ε ( 1 ≤ j ≤ s ) . D_0 \subset D_1 \qquad\text{and}\qquad\|\theta_j|_{D_1'\cap P}\| < \varepsilon\quad(1 \leq j \leq s). D 0 ⊂ D 1 and ∥ θ j ∣ D 1 ′ ∩ P ∥ < ε ( 1 ≤ j ≤ s ) . Proof. Let e 1 , … , e m e_1,\ldots,e_m e 1 , … , e m be the minimal projections of D 0 D_0 D 0 . For each k k k , take a central projection z k ∈ Z ( e k M e k ) z_k \in\mathcal{Z}(e_kMe_k) z k ∈ Z ( e k M e k ) such that
e k M e k = z k ( e k M e k ) ⊕ ( 1 − z k ) ( e k M e k ) , e_kMe_k = z_k(e_kMe_k) \oplus(1-z_k)(e_kMe_k), e k M e k = z k ( e k M e k ) ⊕ ( 1 − z k ) ( e k M e k ) , with z k ( e k M e k ) z_k(e_kMe_k) z k ( e k M e k ) finite and ( 1 − z k ) ( e k M e k ) (1-z_k)(e_kMe_k) ( 1 − z k ) ( e k M e k ) properly infinite. Replacing D 0 D_0 D 0 with D 0 ∨ { z k } ′ ′ D_0 \vee\{z_k\}^{\prime\prime} D 0 ∨ { z k } ′′ and relabelling, we may assume that each e k M e k e_kMe_k e k M e k is either finite or properly infinite. By Lemma 6.1(i), each inclusion e k M e k ⊂ e k P e k e_kMe_k \subset e_kPe_k e k M e k ⊂ e k P e k satisfies the hypotheses of Theorem G , and the restrictions θ j ∣ e k P e k \theta_j|_{e_kPe_k} θ j ∣ e k P e k vanish on e k M e k e_kMe_k e k M e k . For each k k k , apply Proposition 6.6 if e k M e k e_kMe_k e k M e k is finite and Proposition 6.5 if it is properly infinite, with error ε / m \varepsilon/m ε / m . We have a finite-dimensional abelian subalgebra D 1 , k ⊂ e k M e k D_{1,k} \subset e_kMe_k D 1 , k ⊂ e k M e k such that
∥ θ j ∣ D 1 , k ′ ∩ e k P e k ∥ < ε m ( 1 ≤ j ≤ s ) . \|\theta_j|_{D_{1,k}'\cap e_kPe_k}\| < \frac{\varepsilon}{m} \qquad(1 \leq j \leq s). ∥ θ j ∣ D 1 , k ′ ∩ e k P e k ∥ < m ε ( 1 ≤ j ≤ s ) . Let D 1 = ⨁ k = 1 m D 1 , k D_1 = \bigoplus_{k=1}^{m}D_{1,k} D 1 = ⨁ k = 1 m D 1 , k . Then D 0 ⊂ D 1 ⊂ M D_0 \subset D_1 \subset M D 0 ⊂ D 1 ⊂ M . By Lemma 6.1(iii), we have
∥ θ j ∣ D 1 ′ ∩ P ∥ ≤ ∑ k = 1 m ∥ θ j ∣ D 1 , k ′ ∩ e k P e k ∥ < ε ( 1 ≤ j ≤ s ) . \|\theta_j|_{D_1'\cap P}\| \leq\sum_{k=1}^{m}\|\theta_j|_{D_{1,k}'\cap e_kPe_k}\| < \varepsilon\qquad(1 \leq j \leq s). ∥ θ j ∣ D 1 ′ ∩ P ∥ ≤ k = 1 ∑ m ∥ θ j ∣ D 1 , k ′ ∩ e k P e k ∥ < ε ( 1 ≤ j ≤ s ) .
Proof of Theorem G. Write M = ( ⨁ 1 ≤ n ≤ ∞ M I n ) ⊕ M c M = \left(\bigoplus_{1\leq n\leq\infty} M_{I_n}\right) \oplus M_c M = ( ⨁ 1 ≤ n ≤ ∞ M I n ) ⊕ M c , where M I n M_{I_n} M I n is homogeneous of type I n I_n I n and M c M_c M c is continuous. By Lemma 6.1(i), we only need to prove the homogeneous type I case and the continuous case. In the former case, write M = B ( H ) ⊗ ‾ Z ( M ) M = B(\mathcal{H})\overline{\otimes}\mathcal{Z}(M) M = B ( H ) ⊗ Z ( M ) . Let P 0 = Z ( M ) ′ ∩ P P_0 = \mathcal{Z}(M)' \cap P P 0 = Z ( M ) ′ ∩ P . Since B ( H ) ′ ∩ P 0 = M ′ ∩ P = Z ( M ) B(\mathcal{H})' \cap P_0 = M' \cap P = \mathcal{Z}(M) B ( H ) ′ ∩ P 0 = M ′ ∩ P = Z ( M ) , we have
P 0 = B ( H ) ⊗ ‾ ( B ( H ) ′ ∩ P 0 ) = B ( H ) ⊗ ‾ Z ( M ) = M . P_0 = B(\mathcal{H})\overline{\otimes}(B(\mathcal{H})' \cap P_0) = B(\mathcal{H})\overline{\otimes}\mathcal{Z}(M) = M. P 0 = B ( H ) ⊗ ( B ( H ) ′ ∩ P 0 ) = B ( H ) ⊗ Z ( M ) = M .
For any maximal abelian subalgebra A ⊂ M A \subset M A ⊂ M , we then have A ′ ∩ P ⊂ Z ( M ) ′ ∩ P = M A' \cap P \subset\mathcal{Z}(M)' \cap P = M A ′ ∩ P ⊂ Z ( M ) ′ ∩ P = M , and hence A ′ ∩ P = A A' \cap P = A A ′ ∩ P = A . Now assume that M M M is continuous. Since P ∗ P_* P ∗ is separable, take a norm-dense sequence ( θ j ) j ≥ 1 (\theta_j)_{j\geq1} ( θ j ) j ≥ 1 in M ⊥ M_\perp M ⊥ . Starting with D 0 = C 1 D_0 = \mathbb{C}1 D 0 = C 1 and applying Proposition 6.7 to θ 1 , … , θ n \theta_1,\ldots,\theta_n θ 1 , … , θ n with ε = 2 − n \varepsilon= 2^{-n} ε = 2 − n at the n n n th step, we have finite-dimensional abelian subalgebras D n ⊂ M D_n \subset M D n ⊂ M such that
D n − 1 ⊂ D n , ∥ θ j ∣ D n ′ ∩ P ∥ < 2 − n ( 1 ≤ j ≤ n ) . D_{n-1} \subset D_n,\qquad\lVert\theta_j|_{D_n'\cap P}\rVert< 2^{-n}\qquad(1 \leq j \leq n). D n − 1 ⊂ D n , ∥ θ j ∣ D n ′ ∩ P ∥ < 2 − n ( 1 ≤ j ≤ n ) .
Let A 0 = ⋁ n D n A_0 = \bigvee_n D_n A 0 = ⋁ n D n . For every j ≥ 1 j \geq1 j ≥ 1 , we have
∥ θ j ∣ A 0 ′ ∩ P ∥ ≤ ∥ θ j ∣ D n ′ ∩ P ∥ ⟶ 0 ( n → ∞ ) . \lVert\theta_j|_{A_0'\cap P}\rVert\leq\lVert\theta_j|_{D_n'\cap P}\rVert\longrightarrow0\qquad(n \to\infty). ∥ θ j ∣ A 0 ′ ∩ P ∥ ≤ ∥ θ j ∣ D n ′ ∩ P ∥ ⟶ 0 ( n → ∞ ) .
By norm density and the Hahn–Banach theorem, A 0 ′ ∩ P ⊂ M A_0'\cap P \subset M A 0 ′ ∩ P ⊂ M . Let A A A be a maximal abelian subalgebra of M M M containing A 0 A_0 A 0 . Then
A ′ ∩ P ⊂ A 0 ′ ∩ P ⊂ M , A'\cap P \subset A_0'\cap P \subset M, A ′ ∩ P ⊂ A 0 ′ ∩ P ⊂ M ,
and hence A ′ ∩ P = A ′ ∩ M = A A'\cap P = A'\cap M = A A ′ ∩ P = A ′ ∩ M = A , which finishes the proof. □
φ \varphi φ -harmonic ergodicity and AFD ergodicity
Let M ⊂ P M \subset P M ⊂ P be an inclusion of von Neumann algebras with separable preduals such that M ′ ∩ P = Z ( M ) M'\cap P = \mathcal{Z}(M) M ′ ∩ P = Z ( M ) . Following the terminology for ergodic embeddings of Popa [40 Definition 2.2] and the notion of AFD-ergodicity used by Marrakchi [33 ] , we use the following terminology, with Z ( M ) \mathcal{Z}(M) Z ( M ) in place of C 1 \mathbb{C}1 C 1 when M M M is not a factor. We say that M ⊂ P M \subset P M ⊂ P is AFD-ergodic if there is a hyperfinite von Neumann subalgebra N ⊂ M N \subset M N ⊂ M such that N ′ ∩ P = Z ( M ) N'\cap P = \mathcal{Z}(M) N ′ ∩ P = Z ( M ) . We say that M ⊂ P M \subset P M ⊂ P is MV-ergodic if it has the weak Dixmier property. Clearly, AFD-ergodicity implies MV-ergodicity by averaging over the unitary groups of an increasing sequence of finite-dimensional subalgebras generating N N N . We also use Poisson boundaries to define the following notion of ergodicity.
Definition 6.8. Let M ⊂ P M \subset P M ⊂ P be an inclusion of von Neumann algebras with separable preduals, and let φ \varphi φ be an nrsg M M M -hyperstate with Poisson boundary B φ B_\varphi B φ . We say that M ⊂ P M \subset P M ⊂ P has φ \varphi φ -harmonic ergodicity if there is a normal faithful ucp map Φ : P → B φ \Phi: P \to B_\varphi Φ : P → B φ with Φ ∣ M = id M \Phi|_M = \operatorname{id}_M Φ ∣ M = id M .
By Theorem 4.2 (i), we have already proved that φ \varphi φ -harmonic ergodicity implies MV-ergodicity when M M M is properly infinite.
In the case M ′ ∩ P = Z ( M ) M'\cap P = \mathcal{Z}(M) M ′ ∩ P = Z ( M ) , Popa [38 Theorem 3.2 and Remark 3.5(2)] proved that, when M M M is a semifinite factor, the existence of a normal faithful conditional expectation E : P → M E : P \to M E : P → M implies AFD-ergodicity. Marrakchi [33 Theorem B] proved this implication for general M M M . As explained in Subsection 1.3, this is their method for constructing finite-dimensional subalgebras F n ⊂ M F_n \subset M F n ⊂ M with ∥ θ ∣ F n ′ ∩ P ∥ → 0 \lVert\theta|_{F_n'\cap P}\rVert\to0 ∥ θ ∣ F n ′ ∩ P ∥ → 0 for every θ ∈ Z ( M ) ⊥ \theta\in\mathcal{Z}(M)^\perp θ ∈ Z ( M ) ⊥ .
We strengthen these results when M M M is properly infinite by proving that φ \varphi φ -harmonic ergodicity directly implies AFD-ergodicity. We first strengthen Proposition 6.3 by passing from a single full matrix algebra F ⊂ M F \subset M F ⊂ M with ∥ θ j ∣ F ′ ∩ P ∥ < ε \lVert\theta_j|_{F'\cap P}\rVert< \varepsilon ∥ θ j ∣ F ′ ∩ P ∥ < ε to an increasing sequence.
Proposition 6.9. Let M ⊂ P M \subset P M ⊂ P be an inclusion of von Neumann algebras with separable preduals, and assume that M M M is properly infinite. Let φ \varphi φ be an nrsg M M M -hyperstate, and assume that M ⊂ P M \subset P M ⊂ P has φ \varphi φ -harmonic ergodicity. Write ( M ′ ∩ P ) ⊥ = { θ ∈ P ∗ ∣ θ ∣ M ′ ∩ P = 0 } (M'\cap P)^\perp= \{\theta\in P_* \mid\theta|_{M'\cap P} = 0\} ( M ′ ∩ P ) ⊥ = { θ ∈ P ∗ ∣ θ ∣ M ′ ∩ P = 0 } . For every full matrix algebra F 0 ⊂ M F_0 \subset M F 0 ⊂ M , every θ 1 , … , θ s ∈ ( M ′ ∩ P ) ⊥ \theta_1,\ldots,\theta_s \in(M'\cap P)^\perp θ 1 , … , θ s ∈ ( M ′ ∩ P ) ⊥ and every ε > 0 \varepsilon> 0 ε > 0 , there is a full matrix algebra F ⊂ M F \subset M F ⊂ M such that
F 0 ⊂ F and ∥ θ j ∣ F ′ ∩ P ∥ < ε ( 1 ≤ j ≤ s ) . F_0 \subset F \qquad\text{and}\qquad\lVert\theta_j|_{F'\cap P}\rVert< \varepsilon\qquad(1 \leq j \leq s). F 0 ⊂ F and ∥ θ j ∣ F ′ ∩ P ∥ < ε ( 1 ≤ j ≤ s ) .
Proof. Let Q : P → P Q : P \to P Q : P → P be the normal M M M -inner ucp map Q ( x ) = ∑ k g k x g k ∗ Q(x) = \sum_k g_kxg_k^* Q ( x ) = ∑ k g k x g k ∗ associated with φ \varphi φ , as in (55) . By Proposition 3.13, we have Har ( Q , P ) = M ′ ∩ P \operatorname{Har}(Q,P)=M'\cap P Har ( Q , P ) = M ′ ∩ P . The proofs of Lemma 6.2 and Proposition 6.3 use the conditional expectation only to establish this identity. Hence the same arguments apply to functionals θ \theta θ with θ ∣ M ′ ∩ P = 0 \theta|_{M'\cap P}=0 θ ∣ M ′ ∩ P = 0 . For each n ≥ 1 n \geq1 n ≥ 1 , we apply the matrix construction of Proposition 6.3 with error 2 − n 2^{-n} 2 − n to the first n n n elements of a norm-dense sequence in ( M ′ ∩ P ) ⊥ (M'\cap P)^\perp ( M ′ ∩ P ) ⊥ . By norm density, the resulting full matrix algebras F n ⊂ M F_n \subset M F n ⊂ M satisfy
∥ θ ∣ F n ′ ∩ P ∥ ⟶ 0 ( n → ∞ ) \lVert\theta|_{F_n'\cap P}\rVert\longrightarrow0\qquad(n \to\infty) ∥ θ ∣ F n ′ ∩ P ∥ ⟶ 0 ( n → ∞ )
for every θ ∈ ( M ′ ∩ P ) ⊥ \theta\in(M'\cap P)^\perp θ ∈ ( M ′ ∩ P ) ⊥ . Let M 0 = F 0 ′ ∩ M M_{0}=F_{0}^{\prime}\cap M M 0 = F 0 ′ ∩ M and P 0 = F 0 ′ ∩ P P_{0}=F_{0}^{\prime}\cap P P 0 = F 0 ′ ∩ P , so that M = F 0 ⊗ ‾ M 0 M=F_{0}\overline{\otimes}M_{0} M = F 0 ⊗ M 0 and P = F 0 ⊗ ‾ P 0 P=F_{0}\overline{\otimes}P_{0} P = F 0 ⊗ P 0 . Take matrix units ( e i j ) 1 ≤ i , j ≤ d (e_{ij})_{1\leq i,j\leq d} ( e ij ) 1 ≤ i , j ≤ d for F 0 F_{0} F 0 . Since M M M is properly infinite, e 11 ∼ 1 e_{11}\sim1 e 11 ∼ 1 . Take an isometry v ∈ M v\in M v ∈ M with v v ∗ = e 11 vv^{*}=e_{11} v v ∗ = e 11 . Then
π : P → ∼ P 0 with π ( x ) = ∑ i = 1 d e i 1 v x v ∗ e 1 i \pi:P\xrightarrow{\sim}P_{0}\qquad\text{with }\pi(x)=\sum_{i=1}^{d}e_{i1}vxv^{*}e_{1i} π : P ∼ P 0 with π ( x ) = i = 1 ∑ d e i 1 v x v ∗ e 1 i
is a normal ∗ * ∗ -isomorphism with π ( M ) = M 0 \pi(M)=M_{0} π ( M ) = M 0 . Since v , e i j ∈ M v,e_{ij}\in M v , e ij ∈ M , we have π ( z ) = z \pi(z)=z π ( z ) = z for every z ∈ M ′ ∩ P z\in M^{\prime}\cap P z ∈ M ′ ∩ P . Hence θ j ∘ π ∈ ( M ′ ∩ P ) ⊥ \theta_{j}\circ\pi\in(M^{\prime}\cap P)_{\perp} θ j ∘ π ∈ ( M ′ ∩ P ) ⊥ . Take n n n sufficiently large that ∥ ( θ j ∘ π ) ∣ F n ′ ∩ P ∥ < ε \|(\theta_{j}\circ\pi)|_{F_{n}^{\prime}\cap P}\|<\varepsilon ∥ ( θ j ∘ π ) ∣ F n ′ ∩ P ∥ < ε for 1 ≤ j ≤ s 1\leq j\leq s 1 ≤ j ≤ s , and let F = F 0 ∨ π ( F n ) ≅ F 0 ⊗ ‾ π ( F n ) F=F_{0}\vee\pi(F_{n})\cong F_{0}\overline{\otimes}\pi(F_{n}) F = F 0 ∨ π ( F n ) ≅ F 0 ⊗ π ( F n ) . Then F F F is a full matrix algebra containing F 0 F_{0} F 0 , and F ′ ∩ P = π ( F n ′ ∩ P ) F^{\prime}\cap P=\pi(F_{n}^{\prime}\cap P) F ′ ∩ P = π ( F n ′ ∩ P ) . Therefore,
∥ θ j ∣ F ′ ∩ P ∥ = ∥ ( θ j ∘ π ) ∣ F n ′ ∩ P ∥ < ε ( 1 ≤ j ≤ s ) , \|\theta_{j}|_{F^{\prime}\cap P}\|=\|(\theta_{j}\circ\pi)|_{F_{n}^{\prime}\cap P}\|<\varepsilon\qquad(1\leq j\leq s), ∥ θ j ∣ F ′ ∩ P ∥ = ∥ ( θ j ∘ π ) ∣ F n ′ ∩ P ∥ < ε ( 1 ≤ j ≤ s ) ,
which finishes the proof. □
Theorem 6.10 (Theorem H). Let M ⊂ P M\subset P M ⊂ P be an inclusion of von Neumann algebras with separable preduals, and assume that M M M is properly infinite. Let φ \varphi φ be an nrsg M M M -hyperstate. If there is a normal faithful ucp map Φ : P → B φ \Phi:P\to B_{\varphi} Φ : P → B φ with Φ ∣ M = id M \Phi|_{M}=\operatorname{id}_{M} Φ ∣ M = id M , then there is a hyperfinite von Neumann subalgebra N ⊂ M N\subset M N ⊂ M such that N ′ ∩ P = M ′ ∩ P N^{\prime}\cap P=M^{\prime}\cap P N ′ ∩ P = M ′ ∩ P .
Proof. Take a norm-dense sequence ( θ j ) j ≥ 1 (\theta_{j})_{j\geq1} ( θ j ) j ≥ 1 in ( M ′ ∩ P ) ⊥ (M^{\prime}\cap P)_{\perp} ( M ′ ∩ P ) ⊥ . Starting with F 0 = C 1 F_{0}=\mathbb{C}1 F 0 = C 1 , apply Proposition 6.9 recursively. We have full matrix algebras F n ⊂ M F_{n}\subset M F n ⊂ M such that F n − 1 ⊂ F n F_{n-1}\subset F_{n} F n − 1 ⊂ F n ,
∥ θ j ∣ F n ′ ∩ P ∥ < 2 − n ( 1 ≤ j ≤ n ) . \|\theta_{j}|_{F_{n}^{\prime}\cap P}\|<2^{-n}\qquad(1\leq j\leq n). ∥ θ j ∣ F n ′ ∩ P ∥ < 2 − n ( 1 ≤ j ≤ n ) .
Let N = ⋁ n F n N=\bigvee_{n}F_{n} N = ⋁ n F n . Then N N N is hyperfinite. For every j ≥ 1 j\geq1 j ≥ 1 , we have
∥ θ j ∣ N ′ ∩ P ∥ ≤ ∥ θ j ∣ F n ′ ∩ P ∥ ⟶ 0 ( n → ∞ ) . \|\theta_{j}|_{N^{\prime}\cap P}\|\leq\|\theta_{j}|_{F_{n}^{\prime}\cap P}\|\longrightarrow0\qquad(n\to\infty). ∥ θ j ∣ N ′ ∩ P ∥ ≤ ∥ θ j ∣ F n ′ ∩ P ∥ ⟶ 0 ( n → ∞ ) .
By norm density and the Hahn–Banach theorem, N ′ ∩ P ⊂ M ′ ∩ P N^{\prime}\cap P\subset M^{\prime}\cap P N ′ ∩ P ⊂ M ′ ∩ P . Since N ⊂ M N\subset M N ⊂ M , the reverse inclusion is immediate. Hence N ′ ∩ P = M ′ ∩ P N^{\prime}\cap P=M^{\prime}\cap P N ′ ∩ P = M ′ ∩ P . □ Mathematical Institute, University of Oxford, Oxford, United Kingdom
Email address: shuoxing.zhou@maths.ox.ac.uk