Open MathsAll courses
Positive maps and finite-dimensional approximation · Prerequisite proofs · Sources and terms

Normal automorphism transport and Bernoulli fixed algebras

Programme exposition: CC0-1.0 to the extent of rights held. Human-source terms remain separate.

The centrally trivial automorphism lesson uses general factors with separable predual, an arbitrary finite-factor displacement argument, and two different Bernoulli index groups. The finite group supplies noncommuting outer classes. The infinite group supplies decreasing fixed algebras with scalar intersection. We prove the needed constructions at those full scopes.

The complete earlier inputs are CF, CP and BD, NT1–2 and the dyadic model FG2, PC/FCT comparison and the normal finite trace, AF1–5, FH0's normality and uniqueness of finite-factor tracial states, and SS0–3's exact coordinates, whole variable product and spatial splitting. Inner products are linear in the second variable.

CT0. Normal transport, countable tests and the unitary action

Let γ\gamma be a normal automorphism of a von Neumann algebra MM. Precomposition preserves M∗M_*, its norm and positivity. For x,z∈Mx,z\in M,

[γ(x),ψ](γ(z))=[x,ψγ](z),[x,ψ](z)=ψ(zx)−ψ(xz).(CT0.1)[\gamma(x),\psi](\gamma(z))=[x,\psi\gamma](z), \qquad [x,\psi](z)=\psi(zx)-\psi(xz). \tag{CT0.1}

Since γ\gamma bijects the unit ball, the commutator norms agree. Thus γ\gamma and its inverse preserve bounded centralizing sequences, ordinarily or along any free ultrafilter. For each positive normal functional, its two square tests on γ(x)\gamma(x) are the corresponding tests on xx with functional precomposed by γ\gamma. This proves preservation of bounded intrinsic strong* convergence. NT2 converts them to a faithful-state test whenever that state is used.

For M∗M_* separable, choose a countable norm-dense sequence in its unit ball. Pointwise norm convergence on this family is equivalent to pointwise norm convergence on every normal functional, because every precomposition operator is an isometry. A metric formed by the sum of 2−j2^{-j} times each test distance truncated at one induces this topology. Its rational-radius balls give a countable neighborhood basis at each automorphism. The dense-state family may also be used: NT1's four positive normal functionals and normalization to states give the same dense linear span.

Group composition and inversion are continuous in this topology. Indeed the precomposition maps are surjective isometries of the Banach space M∗M_*. For isometries An→AA_n\to A, Bn→BB_n\to B pointwise,

∥AnBnξ−ABξ∥≤∥Bnξ−Bξ∥+∥(An−A)Bξ∥→0.\|A_nB_n\xi-AB\xi\| \leq\|B_n\xi-B\xi\|+\|(A_n-A)B\xi\|\to0.

Also ∥An−1ξ−A−1ξ∥=∥ξ−AnA−1ξ∥→0\|A_n^{-1}\xi-A^{-1}\xi\|=\|\xi-A_nA^{-1}\xi\|\to0. These estimates apply to nets as well. Normality of the subgroup of inner automorphisms follows from γAd⁡(u)γ−1=Ad⁡(γ(u))\gamma\operatorname{Ad}(u)\gamma^{-1}=\operatorname{Ad}(\gamma(u)); continuity then makes its closure a normal subgroup. Its abstract quotient by the inner subgroup is consequently defined even when the inner subgroup is not closed.

For a unitary uu, fixed multiplication by u∗u^* bijects the operator unit ball and gives

∥ψAd⁡(u)−ψ∥=∥[u,ψ]∥.(CT0.2)\|\psi\operatorname{Ad}(u)-\psi\|=\|[u,\psi]\|. \tag{CT0.2}

If un→uu_n\to u strong*, then their inner automorphisms converge in this topology. CP expresses normal functionals as norm-summable vector-pair series. On each pair, conjugation replaces the vectors by un∗ξ,un∗ηu_n^*\xi,u_n^*\eta, which converge in Hilbert norm. The pairing difference on the unit ball is bounded by the changed-vector norm times the other vector norm. Finite truncation and the uniform precomposition norm give convergence of every normal functional. This is the inner-implementation continuity used with the already complete strong* unitary metric.

Conversely, uu-convergence βi→β\beta_i\to\beta of automorphisms gives pointwise strong* convergence. For fixed xx, expand the square of βi(x)−β(x)\beta_i(x)-\beta(x) under a positive normal φ\varphi. Its first square term tests βi(x∗x)\beta_i(x^*x); the two cross terms test βi(x)\beta_i(x) by fixed normal multiplier functionals. Every term converges by the predual topology to its β\beta-value, so the whole square tends to zero. Apply the same argument to x∗x^* for the other square. This argument does not assert a continuity property of the map from automorphisms to an ultrafilter quotient.

Finally γ\gamma induces an automorphism on MωM_\omega whenever the complete quotient construction applies: CT0.1 preserves CωC_\omega, the two square tests preserve JωJ_\omega, and the inverse gives bijectivity of the induced star map. Scalar ultraweak limits are preserved by normality and its predual tests. Hence the quotient trace is preserved. SS1 proves whole-domain normality of this induced map and its inverse.

CT1. Finite matrix implementation and every automorphism of RR

In a finite factor, FH0 proves that any normalized tracial state is the canonical faithful normal trace τ\tau. Thus every normal automorphism α\alpha preserves it: τα\tau\alpha is another normalized tracial state. This uses neither separability nor an additional normality assumption on a trace originally specified only as tracial.

For a unital D≅MdD\cong M_d with units eije_{ij}, let fij=α(eij)f_{ij}=\alpha(e_{ij}). Each family has diagonal trace 1/d1/d, since the diagonals are equivalent and sum to one. FCT7 therefore gives vv with v∗v=e11v^*v=e_{11}, vv∗=f11vv^*=f_{11}. The exact finite sum

w=∑ifi1ve1i(CT1.1)w=\sum_i f_{i1}ve_{1i} \tag{CT1.1}

has w∗w=∑ieii=1w^*w=\sum_i e_{ii}=1 and ww∗=∑ifii=1ww^*=\sum_i f_{ii}=1. Multiplying its indices gives weijw∗=fijwe_{ij}w^*=f_{ij}. This is a complete implementation on the whole DD.

Now use the complete dyadic RR, its increasing DnD_n, and the normal trace projections Pn:R→DnP_n:R\to D_n from AF1. Their trace-vector orthogonal projections give ∥Pn(x)−x∥2→0\|P_n(x)-x\|_2\to0, with ∥Pn(x)∥≤∥x∥\|P_n(x)\|\leq\|x\|. Choose wnw_n by CT1.1 for α∣Dn\alpha|_{D_n}, and let γn=Ad⁡(wn)\gamma_n=\operatorname{Ad}(w_n). For n≥kn\geq k,

∥γn(x)−α(x)∥2≤2∥x−Pk(x)∥2,\|\gamma_n(x)-\alpha(x)\|_2 \leq2\|x-P_k(x)\|_2,

because they agree on DkD_k and both preserve the trace norm. Thus γn(x)→α(x)\gamma_n(x)\to\alpha(x) in trace norm for every xx. Their inverses also converge at every xx:

∥γn−1(x)−α−1(x)∥2=∥x−γn(α−1(x))∥2→0.(CT1.2)\|\gamma_n^{-1}(x)-\alpha^{-1}(x)\|_2 =\|x-\gamma_n(\alpha^{-1}(x))\|_2\to0. \tag{CT1.2}

To prove the full uu-convergence, not only bounded-operator convergence, take the normal functionals ρa,b(x)=τ(a∗xb)\rho_{a,b}(x)=\tau(a^*xb), a,b∈Ra,b\in R. CP's faithful normal trace representation makes their span predual norm dense: truncate any vector-pair series and approximate its vectors by bounded trace vectors. Trace preservation gives

ρa,bγn(x)=τ(γn−1(a)∗xγn−1(b)).\rho_{a,b}\gamma_n(x) =\tau(\gamma_n^{-1}(a)^*x\gamma_n^{-1}(b)).

On ∥x∥≤1\|x\|\leq1, the difference from ρa,bα\rho_{a,b}\alpha is at most

∥γn−1(a)−α−1(a)∥2 ∥b∥2+∥a∥2 ∥γn−1(b)−α−1(b)∥2.\|\gamma_n^{-1}(a)-\alpha^{-1}(a)\|_2\,\|b\|_2 +\|a\|_2\,\|\gamma_n^{-1}(b)-\alpha^{-1}(b)\|_2.

Scalar trace Schwarz proves this bound. CT1.2 makes it vanish; predual density and isometry extend it to every normal functional. Therefore every automorphism of RR is approximately inner.

CT2. The tensor relative commutant and topology transport

Let NN be any nonzero factor, normally represented on HH, and let RR be faithfully normally represented on KK. In the concrete spatial product,

(N⊗1)′∩(N⊗ˉR)=1⊗R.(CT2.1)(N\otimes1)'\cap(N\bar\otimes R)=1\otimes R. \tag{CT2.1}

Here is a direct whole-operator proof. For X∈N⊗ˉRX\in N\bar\otimes R, compression between fixed second-leg vectors ξ,η\xi,\eta gives an operator Xξ,ηX_{\xi,\eta} on HH. It commutes with N′N', since XX commutes with N′⊗1N'\otimes1; hence Xξ,η∈N′′=NX_{\xi,\eta}\in N''=N. If XX also commutes with N⊗1N\otimes1, every such coefficient lies in Z(N)=CZ(N)=\mathbb C.

Choose a unit first-leg vector h0h_0. Compression between h0,h0h_0,h_0 defines a bounded T∈B(K)T\in B(K), with ∥T∥≤∥X∥\|T\|\leq\|X\|. The scalar coefficient identities give every elementary matrix coefficient of XX equal to that of 1⊗T1\otimes T; density of finite tensor vectors proves equality of the whole operators. Since XX commutes with 1⊗R′1\otimes R', T∈R′′=RT\in R''=R. The converse inclusion is direct. SS3 supplies the complete Hilbert tensor construction used for these vector compressions.

For topology transport assume NN has a faithful normal state, as it does in the general-factor application with separable predual. If γn→γ\gamma_n\to\gamma in the uu-topology on RR, then id⁡N⊗γn→id⁡N⊗γ\operatorname{id}_N\otimes\gamma_n\to\operatorname{id}_N\otimes\gamma in that topology on N⊗ˉRN\bar\otimes R. The normal tensor maps exist on the whole product: represent NN by a faithful normal state and RR by its preserved faithful normal trace, apply the trace-GNS unitary implementing γ\gamma on the second leg and conjugate. Its normal inverse follows from CP and the faithful normal GNS identifications. The general-factor application has separable predual and such a faithful normal state, as NT2 proves.

For the convergence assertion, finite sums of elementary normal tensor functionals are predual norm dense by CP: truncate a vector-pair series and approximate each vector by finite tensors. Their elementary norm bound is ∥η⊗ζ∥≤∥η∥∥ζ∥\|\eta\otimes\zeta\|\leq\|\eta\|\|\zeta\|, proved by taking almost norm-minimizing trace-class extensions to each represented B(H)B(H) and tensoring their summable vector-pair series, then letting the norm errors tend to zero. Thus

∥(η⊗ζ)(id⁡⊗γn)−(η⊗ζ)(id⁡⊗γ)∥≤∥η∥ ∥ζγn−ζγ∥→0.\|(\eta\otimes\zeta)(\operatorname{id}\otimes\gamma_n) -(\eta\otimes\zeta)(\operatorname{id}\otimes\gamma)\| \leq\|\eta\|\,\|\zeta\gamma_n-\zeta\gamma\|\to0.

The uniform precomposition norm makes the same conclusion hold on the entire predual. If γn=Ad⁡(wn)\gamma_n=\operatorname{Ad}(w_n), their tensor maps are implemented by 1⊗wn1\otimes w_n.

CT3. Bernoulli products for finite and countably infinite groups

Let K0K_0 be a nonempty finite or countably infinite discrete group. For each h∈K0h\in K_0 take a dyadic RhR_h, with the same identified finite heads DrD_r. Construct the complete product BB by placing one M2M_2 at every index in K0×NK_0\times\mathbb N. This index set is countably infinite even when K0K_0 is finite. SS2 constructs its faithful normal product trace representation and identifies its whole closure with RR.

Grouping the legs at a fixed hh gives Rh⊆BR_h\subseteq B. On local trace vectors every regrouping or countable permutation preserves all finite product inner products. It consequently extends to a Hilbert unitary, intertwining the local actions and their BD closures normally in both directions. This identifies BB with the complete product of the RhR_h's. A finite set F⊆K0F\subseteq K_0 supports its corresponding whole finite product. Trace factorization for operators on disjoint supports follows first on finite matrix legs and then by normality in each factor on the whole supported algebras.

The left Bernoulli convention is

αg(⨂hxh)=⨂hxg−1h.(CT3.1)\alpha_g\left(\bigotimes_h x_h\right) =\bigotimes_h x_{g^{-1}h}. \tag{CT3.1}

It sends an operator formerly in coordinate hh to coordinate ghgh. On finite products its compositions satisfy αgαk=αgk\alpha_g\alpha_k=\alpha_{gk}, preserve multiplication, adjoints and the product trace, and have inverse αg−1\alpha_{g^{-1}}. The trace-vector unitary extends them to the whole BB; CP's unitary-conjugation tests and the faithful normal GNS inverse prove normality. The group identities extend by BD.

In coordinate ee, choose ar=e11−e22a_r=e_{11}-e_{22} at its rr-th dyadic leg, with identities at the other legs. These are trace-zero self-adjoint unitaries. Write a~r\widetilde a_r for their images in BB. Each fixed finite matrix stage commutes with them eventually. For x∈Bx\in B, bounded finite-stage trace approximation therefore gives ∥[a~r,x]∥2≤2∥x−x0∥2\|[\widetilde a_r,x]\|_2\leq2\|x-x_0\|_2 at all sufficiently large rr. Adjoint estimates are the same; their trace functional commutator is zero. The complete tracial centrality criterion makes this an ordinary centralizing sequence.

For g≠eg\ne e, αg(a~r)\alpha_g(\widetilde a_r) has distinct coordinate support gg. Both unitaries have trace zero and trace square norm one. Product trace factorization yields

∥αg(a~r)−a~r∥22=2.(CT3.2)\|\alpha_g(\widetilde a_r)-\widetilde a_r\|_2^2=2. \tag{CT3.2}

Every inner automorphism moves a bounded centralizing sequence by a strong*-zero difference, hence a trace-norm-zero difference here. CT3.2 rules this out. Thus each nonidentity αg\alpha_g is outer, for finite groups as well as infinite ones. In particular K0=S3K_0=S_3 supplies a faithful everywhere-outer action on an algebra normally isomorphic to RR.

CT4. Direct irreducibility and type of the finite fixed algebras

Now let K0K_0 be countably infinite and H0⊆K0H_0\subseteq K_0 a finite subgroup, m=∣H0∣m=|H_0|. Set N=BH0N=B^{H_0}. It is a von Neumann subalgebra, since each fixed equation is ultraweakly closed. The finite map

EH0(x)=m−1∑h∈H0αh(x)E_{H_0}(x)=m^{-1}\sum_{h\in H_0}\alpha_h(x)

is normal, unital, completely positive, trace-preserving, NN-bimodular and has exactly range NN. It is a trace 22-contraction by the finite triangle inequality.

We prove N′∩B=CN'\cap B=\mathbb C without a crossed-product assertion. Fix x∈N′∩Bx\in N'\cap B and a finite-support operator aa. Choose a finite matrix-stage xLx_L, supported on finite LL, with ∥x−xL∥2<δ\|x-x_L\|_2<\delta. Let FF support aa. Infinitude permits a k∉H0(L∪F)k\notin H_0(L\cup F); thus all distinct coordinates hkhk, h∈H0h\in H_0, lie outside L∪H0FL\cup H_0F.

Take a trace-zero unitary bb in coordinate kk, and put bh=αh(b)b_h=\alpha_h(b). The operator

t=∑h∈H0αh(a)bht=\sum_{h\in H_0}\alpha_h(a)b_h

is fixed by H0H_0, since each group element permutes the summands. Its norm is at most m∥a∥m\|a\|, and [x,t]=0[x,t]=0. The bhb_h's commute with xLx_L and with every coefficient [xL,αj(a)][x_L,\alpha_j(a)]. Their mutually disjoint trace-zero supports give τ(bh∗bj)=δhj\tau(b_h^*b_j)=\delta_{hj}. Full product trace factorization therefore proves

∥[xL,t]∥22=∑h∈H0∥[xL,αh(a)]∥22.(CT4.1)\|[x_L,t]\|_2^2 =\sum_{h\in H_0}\|[x_L,\alpha_h(a)]\|_2^2. \tag{CT4.1}

Meanwhile ∥[xL,t]∥2≤2m∥a∥δ\|[x_L,t]\|_2\leq2m\|a\|\delta. The identity-index term in CT4.1 gives

∥[x,a]∥2≤2(m+1)∥a∥δ.\|[x,a]\|_2\leq2(m+1)\|a\|\delta.

Let δ\delta decrease to zero. Hence xx commutes with all finite-support operators, including the generating finite matrix stages. Fixed multiplication and BD then make it commute with BB. The complete product is a factor, so xx is scalar.

Thus NN is a factor, since Z(N)⊆N′∩BZ(N)\subseteq N'\cap B. Its restricted trace is faithful normal, making it finite. It has no minimal projection. In one orbit H0kH_0k, take the diagonal tensor cells of ⨂h∈H0Dr\bigotimes_{h\in H_0}D_r, d=2rd=2^r. Their dmd^m equal trace masses are d−md^{-m}. The sum of the cells in each H0H_0-orbit is a nonzero fixed projection, of trace at most m/dmm/d^m, and these sums partition the unit. If q∈Nq\in N were minimal, choose rr with m/dm<τ(q)m/d^m<\tau(q). FCT7 within the finite factor NN places a nonzero copy of one orbit-sum projection under qq, with strictly smaller trace, a contradiction. PC4/FH1 consequently makes NN type II1\mathrm{II}_1.

For AFD, enumerate K0K_0 by k1,k2,…k_1,k_2,\ldots, put Fr=H0{k1,…,kr}F_r=H_0\{k_1,\ldots,k_r\}, and let

Br=⨂h∈FrDr.B_r=\bigotimes_{h\in F_r}D_r.

These are increasing finite full matrix algebras, invariant under H0H_0, and their union contains every finite matrix leg eventually. AF1's trace projections onto BrB_r give bounded xr∈Brx_r\in B_r with xr→xx_r\to x in trace norm. At x∈Nx\in N, their finite averages belong to the finite-dimensional BrH0B_r^{H_0}, remain bounded and converge in trace norm to xx. NT2 gives bounded strong* convergence; therefore these fixed stages generate NN, proving AFD.

Its predual is separable. For η∈N∗\eta\in N_*, the map η↦ηEH0\eta\mapsto\eta E_{H_0} embeds it isometrically in B∗B_*, because the finite expectation fixes its unit ball. The image is closed, and the complete countable product has separable predual. AF5 now identifies NN normally with RR. The argument includes the trivial subgroup, and uses infinitude only where a remote coordinate is required.

CT5. Compact minimization of a twisted unitary orbit

Let MM be any II1\mathrm{II}_1 factor, without separability, with its canonical normal trace from FH0. For a normal automorphism α\alpha and a unital matrix factor DD, put C=D′∩MC=D'\cap M and d=sup⁡u∈U(C)∥α(u)−u∥2<1d=\sup_{u\in\mathcal U(C)}\|\alpha(u)-u\|_2<1. The orbit uα(u∗)u\alpha(u^*) lies in the operator unit ball and the trace ball of radius dd about 11. Let KK be its ultraweakly closed convex hull.

CP/CF makes the bounded operator ball ultraweakly compact. The scalar formula

∥z∥2=sup⁡a∈M∥a∥2≤1∣τ(a∗z)∣\|z\|_2=\sup_{\substack{a\in M\\\|a\|_2\leq1}}|\tau(a^*z)|

follows from Schwarz and a=z/∥z∥2a=z/\|z\|_2 for nonzero zz. Each pairing is normal; hence centered trace balls are ultraweakly closed. Consequently KK is compact and still lies in the radius-dd ball about 11; it excludes zero.

If s=inf⁡z∈K∥z∥2s=\inf_{z\in K}\|z\|_2, the compact nonempty nested sublevel sets with bounds s+1/ns+1/n have nonempty intersection. They supply a least-norm y∈Ky\in K. If two points had that norm, their midpoint's squared norm would be s2s^2 minus one quarter of their squared distance. Trace faithfulness and convexity therefore prove uniqueness. No Hilbert convex-set projection theorem is needed.

For v∈U(C)v\in\mathcal U(C), the normal linear map z↦vzα(v∗)z\mapsto vz\alpha(v^*) bijects KK and is a trace-norm isometry. It sends the orbit element at uu to the one at vuvu, with inverse the map at v∗v^*. Uniqueness fixes yy. Since 0∉K0\notin K, y≠0y\ne0, and the resulting identity is

yα(v)=vy(v∈U(C)).y\alpha(v)=vy\qquad(v\in\mathcal U(C)).

NT1's complete four-unitary span extends it to all v∈Cv\in C. CT1's matrix implementer and SS0's coefficients finish the innerness argument in the main lesson.