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

Cone geometry, support corners, and positive functionals

Written by GPT-6.1 Sol (OpenAI), Ultra, October 2026. Original text and embedded diagram: public domain (CC0).

Normal positive functionals need not be faithful, and a von Neumann algebra need not admit a faithful normal state. We first control cone vectors and their support projections, then reduce each functional to its own supported corner. The realization theorem needed in that corner is the cyclic theorem stated below.

Constructing the cone and representing functionals

The inner product is linear in its second variable. Thus

ωξ(a)=⟨ξ,aξ⟩. \omega_\xi(a)=\langle\xi,a\xi\rangle.

For a first-variable-linear source inner product (⋅∣⋅)(\cdot\mid\cdot), our convention is ⟨ξ,η⟩=(η∣ξ)\langle\xi,\eta\rangle=(\eta\mid\xi). An antiunitary involution satisfies ⟨Jξ,Jη⟩=⟨η,ξ⟩\langle J\xi,J\eta\rangle=\langle\eta,\xi\rangle.

Let NN be an arbitrary von Neumann algebra. Choose a faithful normal semifinite weight on NN, and use its full left Hilbert algebra to construct a faithful normal representation N≅M⊂B(H)N\cong M\subset B(H). The choice of weight and the faithful full-Hilbert-algebra realization are part of the construction input. In this chapter the form is this constructed form. The modular construction needed here has the following exact output, denoted MC:

  1. JJ is an antiunitary involution, JMJ=M′JMJ=M', and JzJ=z∗JzJ=z^* for z∈Z(M)z\in Z(M).
  2. P⊂HP\subset H is a closed convex cone, fixed pointwise by JJ, and self-dual:
    P={u∈H:⟨u,v⟩∈[0,∞) for every v∈P}.P=\{u\in H:\langle u,v\rangle\in[0,\infty)\text{ for every }v\in P\}.
  3. For every a∈Ma\in M, aJaJ(P)⊂PaJaJ(P)\subset P.
  4. In every cyclic separating representation of a von Neumann algebra A⊂B(K)A\subset B(K), including A=qMq⊂B(qH)A=qMq\subset B(qH) for the corners below, the closure SS of aγ↦a∗γa\gamma\mapsto a^*\gamma has adjoint core A′γA'\gamma, with S∗(b′γ)=b′∗γS^*(b'\gamma)=b'^*\gamma, and has the modular polar decomposition. Its natural cone
    Pγ={aJγaJγγ:a∈A}‾P_\gamma=\overline{\{aJ_\gamma aJ_\gamma\gamma:a\in A\}}
    is self-dual.

The commutant has the same geometric data (H,J,P)(H,J,P). Indeed JM′J=MJM'J=M, its center equals Z(M)Z(M), and, if a′=JaJ∈M′a'=JaJ\in M', then

a′Ja′J=(JaJ)a=aJaJ. a'Ja'J=(JaJ)a=aJaJ.

The two factors commute, so this operator preserves PP by MC3. The cyclic-core output MC4 already applies to every cyclic separating representation. Thus the arguments below apply to the commutant with this same cone.

The construction of MC is proved in Constructing the natural cone from bounded multiplication: use the faithful normal semifinite weight chosen in NC01, whose representation, original involution graph and fullness are constructed in NC01b. NC00 specifies the earlier operator constructions, NC02–NC03 prove positive form extension and endpoint duality, NC04–NC06 construct the self-dual preserved cone, and NC07–NC08 prove the central and universal cyclic identities. The bounded modular route RC–GP–RS–IK–MC–MP supplies the commutant and original polar-operator graph used in NC00.

The cyclic realization statement, denoted CR, is:

In the natural cone of a cyclic separating vector, every bounded normal positive functional has a representing vector in that cone.

Normal positive functionals in a cyclic natural cone proves CR. Its positive Fourier correction, supported-corner domain calculation, residual bound by a factor of 1/41/4, and real separation proof give the representing vector as an actual norm limit. The geometry below proves uniqueness before using CR, and then passes from cyclic corners to arbitrary, possibly nonfaithful functionals. Bounded continuous functional calculus, the bicommutant characterization, and Hilbert-space completeness are the elementary operator prerequisites.

CG01. Normal supports without a representing vector

A bounded increasing net (ai)(a_i) of positive operators has a strong supremum. Indeed, its scalar quadratic forms have limits, polarization gives a bounded positive operator aa, and

∥(a−ai)u∥2≤C⟨u,(a−ai)u⟩⟶0 \|(a-a_i)u\|^2\le C\langle u,(a-a_i)u\rangle\longrightarrow0

when 0≤a−ai≤C10\le a-a_i\le C1. A strongly closed algebra contains this supremum. The same argument applies to increasing nets of projections.

Let ω\omega be a bounded normal positive functional on MM. Positivity gives the Cauchy–Schwarz inequality

∣ω(x∗y)∣2≤ω(x∗x)ω(y∗y). |\omega(x^*y)|^2\le\omega(x^*x)\omega(y^*y).

For completeness, apply positivity to (x+ty)∗(x+ty)(x+ty)^*(x+ty), minimize its scalar quadratic polynomial when ω(y∗y)>0\omega(y^*y)>0, and use arbitrarily small positive denominators when ω(y∗y)=0\omega(y^*y)=0. Consequently a projection ee with ω(e)=0\omega(e)=0 satisfies ω(xe)=ω(ex)=0\omega(xe)=\omega(ex)=0 for every xx.

Null projections are closed under joins, including uncountable joins. If e,fe,f are null and h=e+fh=e+f, then ω(h)=0\omega(h)=0. The continuous-calculus contractions

h(h+ε)−1≤ε−1h h(h+\varepsilon)^{-1}\le\varepsilon^{-1}h

increase strongly to the range projection of hh, which is e∨fe\vee f: ker⁡h=ker⁡e∩ker⁡f\ker h=\ker e\cap\ker f. Normality gives ω(e∨f)=0\omega(e\vee f)=0. The directed net of finite joins of any family of null projections has a strong supremum and zero ω\omega-value. Let rr be the join of all null projections and put

s(ω)=1−r. s(\omega)=1-r.

This is the least projection pp for which ω(1−p)=0\omega(1-p)=0. Cauchy–Schwarz gives

ω(x)=ω(pxp),p=s(ω).(1) \omega(x)=\omega(pxp),\qquad p=s(\omega). \tag{1}

The restriction to pMppMp is faithful. To see this, if b∈(pMp)+b\in(pMp)_+ and ω(b)=0\omega(b)=0, the same contractions b(b+ε)−1b(b+\varepsilon)^{-1} show that the range projection of bb is null. It lies below both pp and rr, so it is zero and b=0b=0.

If ψ≤ω\psi\le\omega, then ψ(1−s(ω))=0\psi(1-s(\omega))=0, and hence

s(ψ)≤s(ω).(2) s(\psi)\le s(\omega). \tag{2}

This is an order statement about functional supports. It makes no assertion that functional order is the order of their eventual cone vectors.

CG02. Algebra and commutant supports of a cone vector

For any ξ∈H\xi\in H, the projection sξs_\xi onto M′ξ‾\overline{M'\xi} belongs to MM: the subspace reduces M′M'. It is the least projection in MM fixing ξ\xi. Since

ωξ(1−e)=∥(1−e)ξ∥2 \omega_\xi(1-e)=\|(1-e)\xi\|^2

for a projection e∈Me\in M, CG01 implies sξ=s(ωξ)s_\xi=s(\omega_\xi). Vector functionals are normal because bounded increasing positive nets converge strongly, as proved above.

If ξ∈P\xi\in P, then Jξ=ξJ\xi=\xi, and applying JJ to M′ξ‾\overline{M'\xi} gives Mξ‾\overline{M\xi}. Therefore the projection qξq_\xi onto the latter subspace satisfies

qξ=JsξJ.(3) \boxed{q_\xi=Js_\xi J.} \tag{3}

In particular ξ\xi is separating for MM exactly when sξ=1s_\xi=1, and is cyclic exactly when qξ=1q_\xi=1. The same conclusions hold on every support corner constructed below.

CG03. Orthogonality detects disjoint supports

For ξ,η∈P\xi,\eta\in P,

⟨η,ξ⟩=0⟺sξsη=0.(4) \langle\eta,\xi\rangle=0 \quad\Longleftrightarrow\quad s_\xi s_\eta=0. \tag{4}

Here is a direct proof of the nontrivial implication. Fix a∈Ma\in M and z∈Cz\in\mathbb C. Cone preservation and self-duality give

0≤⟨η,(1+za)J(1+za)Jξ⟩=2Re⁡(zA)+∣z∣2C,(5) 0\le\langle\eta,(1+za)J(1+za)J\xi\rangle =2\operatorname{Re}(zA)+|z|^2C, \tag{5}

where A=⟨η,aξ⟩A=\langle\eta,a\xi\rangle and C=⟨η,aJaJξ⟩≥0C=\langle\eta,aJaJ\xi\rangle\ge0. In this convention J(1+za)J=1+z‾JaJJ(1+za)J=1+\overline zJaJ, and ⟨η,JaJξ⟩=A‾\langle\eta,JaJ\xi\rangle=\overline A, so the displayed expansion includes both scalar conjugations. If A≠0A\ne0, choose z=−tA‾/∣A∣z=-t\overline A/|A| and let t>0t>0 decrease to zero. The right side is −2t∣A∣+t2C<0-2t|A|+t^2C<0 for small tt, a contradiction. Thus η⊥Mξ\eta\perp M\xi.

By CG02, qξη=0q_\xi\eta=0. Conjugating by JJ gives sξη=0s_\xi\eta=0. The projection 1−sξ1-s_\xi fixes η\eta, so minimality of sηs_\eta gives sη≤1−sξs_\eta\le1-s_\xi. Conversely, disjoint supports put ξ\xi and η\eta in orthogonal projection ranges. Applying JJ also gives qξqη=0q_\xi q_\eta=0.

CG04. Orthogonal positive and negative parts

Let HJ={v:Jv=v}H_J=\{v:Jv=v\}, regarded as a real Hilbert space. Every v∈HJv\in H_J has a unique decomposition

v=v+−v−,v+,v−∈P,⟨v+,v−⟩=0.(6) v=v_+-v_-,\qquad v_+,v_-\in P, \qquad\langle v_+,v_-\rangle=0. \tag{6}

To prove it, let d=inf⁡u∈P∥v−u∥d=\inf_{u\in P}\|v-u\| and choose a minimizing sequence unu_n. The parallelogram identity and (un+um)/2∈P(u_n+u_m)/2\in P give

∥un−um∥2≤2∥v−un∥2+2∥v−um∥2−4d2⟶0. \|u_n-u_m\|^2 \le2\|v-u_n\|^2+2\|v-u_m\|^2-4d^2\longrightarrow0.

Its limit p∈Pp\in P attains the minimum. Varying pp to p+tup+t u, u∈Pu\in P, t≥0t\ge0, gives ⟨p−v,u⟩≥0\langle p-v,u\rangle\ge0. Varying to (1+t)p(1+t)p for small positive and negative tt gives ⟨p−v,p⟩=0\langle p-v,p\rangle=0. Self-duality gives n=p−v∈Pn=p-v\in P, and v=p−nv=p-n with p⊥np\perp n.

If v=p−nv=p-n is any such decomposition, then for u∈Pu\in P,

∥v−u∥2=∥p−u∥2+∥n∥2+2⟨n,u⟩≥∥n∥2. \|v-u\|^2=\|p-u\|^2+\|n\|^2+2\langle n,u\rangle\ge\|n\|^2.

Equality holds at u=pu=p, and can hold only there. Thus pp is the unique closest point and n=p−vn=p-v is unique too. CG03 shows that their algebra supports are disjoint.

Every v∈Hv\in H is a complex linear combination of four cone vectors: apply (6) to (v+Jv)/2(v+Jv)/2 and (v−Jv)/(2i)(v-Jv)/(2i). In particular the complex span of PP is dense in HH, indeed equals HH.

CG05. The two norm bounds and uniqueness

For ξ,η∈P\xi,\eta\in P,

∥ξ−η∥2≤∥ωξ−ωη∥≤∥ξ−η∥ ∥ξ+η∥.(7) \boxed{\|\xi-\eta\|^2\le\|\omega_\xi-\omega_\eta\| \le\|\xi-\eta\|\,\|\xi+\eta\|.} \tag{7}

Write d=ξ−η=p−nd=\xi-\eta=p-n as in CG04, and let e=spe=s_p, f=snf=s_n. These projections are orthogonal. If δ=ωξ−ωη\delta=\omega_\xi-\omega_\eta, then ed=ped=p and fd=−nfd=-n, so

δ(e)=∥p∥2+2⟨p,η⟩≥∥p∥2,δ(f)=−∥n∥2−2⟨n,ξ⟩≤−∥n∥2.\begin{aligned} \delta(e)&=\|p\|^2+2\langle p,\eta\rangle\ge\|p\|^2,\\ \delta(f)&=-\|n\|^2-2\langle n,\xi\rangle\le-\|n\|^2. \end{aligned}

The pairings are real and nonnegative by self-duality. Since e−fe-f is a self-adjoint contraction,

∥δ∥≥δ(e−f)≥∥p∥2+∥n∥2=∥d∥2. \|\delta\|\ge\delta(e-f) \ge\|p\|^2+\|n\|^2=\|d\|^2.

This proves the lower bound without first assuming that any given normal functional has a representative.

For a Hermitian functional the norm can be tested on self-adjoint contractions: for any contraction aa, choose a scalar λ\lambda of modulus one with λδ(a)=∣δ(a)∣\lambda\delta(a)=|\delta(a)|, and test (λa+λ‾a∗)/2(\lambda a+\overline\lambda a^*)/2. For self-adjoint aa,

δ(a)=Re⁡⟨ξ−η,a(ξ+η)⟩. \delta(a)=\operatorname{Re}\langle\xi-\eta,a(\xi+\eta)\rangle.

Cauchy–Schwarz proves the upper bound in (7). For any positive functional ω\omega, Cauchy–Schwarz gives ∣ω(a)∣2≤ω(1)ω(a∗a)≤ω(1)2|\omega(a)|^2\le\omega(1)\omega(a^*a)\le\omega(1)^2 for a contraction aa; testing at one therefore gives ∥ω∥=ω(1)\|\omega\|=\omega(1). In particular ∥ωξ∥=∥ξ∥2\|\omega_\xi\|=\|\xi\|^2. The lower bound proves that a normal positive functional has at most one representative in PP. It also shows that the set of functionals represented in PP is norm closed: a norm-Cauchy sequence of represented functionals has norm-Cauchy representatives, whose limit lies in PP and represents the limit by the upper bound.

Positive decomposition and norm bound in the two-point algebra The positive quadrant is the natural cone for the diagonal algebra C squared. Xi is three comma one, eta is one comma two, their difference is two comma minus one, its positive part is two comma zero, and its negative part is zero comma one. The squared distance is five and the functional difference norm is eleven. The support test controls the vector distance Exact model: M = C², H = C², J = coordinate conjugation, P = R²₊ 0 123 12−1 P ξ = (3, 1) η = (1, 2) d = (2, −1) p = (2, 0) n = (0, 1) −n Orthogonal cone parts (CG04) d = p − n, ⟨p, n⟩ = 0 p is the closest cone point to d. Orthogonal support projections (CG03) e = diag(1, 0), f = diag(0, 1) e f = 0, ‖e − f‖ = 1 The norm test (CG05) ωξ = (9, 1), ωη = (1, 4) δ(e − f) = 8 + 3 = ‖δ‖ = 11 ‖d‖² = 4 + 1 = 5 ≤ 11 Excess = 2⟨p, η⟩ + 2⟨n, ξ⟩ = 6 The diagram depicts this two-point model. The general proof uses a self-dual cone and projection supports. Proofs CG03–CG05; related conclusions: Araki 1974, Theorem 4(6)–(8), pp. 328–332; Haagerup 1975, Lemma 2.10.

In this figure M=C2M=\mathbb C^2 acts diagonally on H=C2H=\mathbb C^2, JJ is coordinate conjugation, and P=R+2P=\mathbb R_+^2. With ξ=(3,1)\xi=(3,1) and η=(1,2)\eta=(1,2), the closest cone point to d=(2,−1)d=(2,-1) is p=(2,0)p=(2,0), and n=(0,1)n=(0,1). The support test e−f=diag⁡(1,−1)e-f=\operatorname{diag}(1,-1) gives ∥ωξ−ωη∥=∣8∣+∣−3∣=11≥5=∥d∥2\|\omega_\xi-\omega_\eta\|=|8|+|-3|=11\ge5=\|d\|^2. The excess is 2⟨p,η⟩+2⟨n,ξ⟩=62\langle p,\eta\rangle+2\langle n,\xi\rangle=6. This is a concrete commutative example of CG04–CG05, not a claim that a general natural cone is two-dimensional. Araki's related general conclusions are Theorem 4(6)–(8), printed pages 328–332; Haagerup's two bounds are Lemma 2.10, printed pages 278–279.

CG06. A corner commutant proved by a Gram form

We need a corner theorem at arbitrary Hilbert-space cardinality. If A⊂B(K)A\subset B(K) is a von Neumann algebra and p∈Ap\in A is a projection, then, on pKpK,

(pAp)′=pA′p.(8) (pAp)'=pA'p. \tag{8}

Here is its full proof. Let T∈(pAp)′T\in(pAp)', acting on pKpK. On the algebraic span of ApKApK define

T^(∑iaiui)=∑iaiTui,ai∈A,ui∈pK.(9) \widehat T\Big(\sum_i a_i u_i\Big)=\sum_i a_i T u_i, \qquad a_i\in A,\quad u_i\in pK. \tag{9}

The positive matrix G=(pai∗ajp)ijG=(pa_i^*a_jp)_{ij} on (pK)n(pK)^n gives ∥∑iaiui∥2=⟨u,Gu⟩\|\sum_i a_i u_i\|^2=\langle u,Gu\rangle. The diagonal operator D=diag⁡(T,…,T)D=\operatorname{diag}(T,\ldots,T) and its adjoint commute with GG, because its entries lie in pAppAp. Continuous functional calculus gives commutation with G1/2G^{1/2}; hence

∥∑iaiTui∥2=∥G1/2Du∥2≤∥T∥2∥G1/2u∥2. \Big\|\sum_i a_i T u_i\Big\|^2 =\|G^{1/2}Du\|^2 \le\|T\|^2\|G^{1/2}u\|^2.

This proves that (9) is well defined and bounded, even when the displayed vector has more than one expression.

The closed subspace L=ApK‾L=\overline{ApK} reduces both AA and A′A'. Thus its projection is central in AA. Extend T^\widehat T by zero on L⊥L^\perp. On the dense algebraic span it commutes with every member of AA, and therefore it belongs to A′A'. Its restriction to pKpK is TT. This gives the difficult inclusion in (8); the other inclusion follows by multiplication. The proof uses finite Gram matrices, not a countable generating set.

Let zA(p)z_A(p) be the projection onto ApK‾\overline{ApK}. The preceding reduction shows that it is the least central projection above pp. If e∈Ae\in A, f∈A′f\in A' are projections and ef=0ef=0, then ff kills AeKAeK, so fzA(e)=0f z_A(e)=0. Taking the least central projection above ff gives

zA(e)zA′(f)=0.(10) z_A(e)z_{A'}(f)=0. \tag{10}

Finally,

Z(pAp)=pZ(A).(11) Z(pAp)=pZ(A). \tag{11}

To see this, extend T∈Z(pAp)T\in Z(pAp) by (9). The operator TT itself belongs to AA, so it commutes with every b∈A′b\in A' on pKpK. On ApKApK, equation (9) then shows T^b=bT^\widehat T b=b\widehat T. Both extensions preserve LL, and the extension is zero on L⊥L^\perp, so T^\widehat T also commutes with A′A'. The bicommutant theorem puts it in A∩A′=Z(A)A\cap A'=Z(A), with pT^p=Tp\widehat T p=T. The reverse inclusion in (11) is immediate.

CG07. The simultaneous support corner

For p∈Mp\in M put j(p)=JpJj(p)=JpJ and q=pj(p)q=pj(p). The central identity in MC gives

zM′(j(p))=zM(p). z_{M'}(j(p))=z_M(p).

If p≠0p\ne0, then q≠0q\ne0, since otherwise (10) would make this nonzero central projection orthogonal to itself.

The map

θ:pMp⟶qMq⊂B(qH),θ(a)=a∣qH,(12) \theta:pMp\longrightarrow qMq\subset B(qH), \qquad \theta(a)=a|_{qH}, \tag{12}

is a faithful normal unital *-isomorphism, and

(qMq)′=qM′qon qH.(13) (qMq)'=qM'q\quad\text{on }qH. \tag{13}

To verify (13) without an induction theorem, first apply (8) to M,pM,p on HH, obtaining (pMp)′=pM′p(pMp)'=pM'p on pHpH. There qq belongs to the latter commutant. Apply (8) again, now to pM′ppM'p and its projection qq. It gives (qM′q)′=qMq(qM'q)'=qMq. In particular qMqqMq is a von Neumann algebra; taking commutants gives (13).

Because j(p)j(p) commutes with pMppMp, (12) is multiplicative. It is onto because qxq=q(pxp)qqxq=q(pxp)q for x∈Mx\in M. For injectivity suppose a∈pMpa\in pMp and aq=0aq=0. Then aj(p)=0aj(p)=0. Let e≤pe\le p be the range projection of ∣a∣|a|, obtained by increasing continuous-calculus contractions. Commutation with j(p)j(p) gives ej(p)=0ej(p)=0. By (10), zM(e)z_M(e) is orthogonal to zM′(j(p))=zM(p)z_{M'}(j(p))=z_M(p), while zM(e)≤zM(p)z_M(e)\le z_M(p). Hence e=0e=0 and a=0a=0.

A faithful *-isomorphism reflects positivity: apply it to the negative part of a self-adjoint operator, using continuous functional calculus. It is therefore an order isomorphism. Suprema of bounded increasing positive nets are characterized by order alone, so both θ\theta and its inverse preserve them. This proves the asserted normality.

The restriction Jq=J∣qHJ_q=J|_{qH} is an antiunitary involution, since JqJ=qJqJ=q. The cone

Pq=qP=P∩qH(14) P_q=qP=P\cap qH \tag{14}

is closed: qP⊂PqP\subset P follows from MC with a=pa=p, and a vector of P∩qHP\cap qH is its own image under qq. If v∈qHv\in qH pairs nonnegatively with PqP_q, then ⟨v,u⟩=⟨v,qu⟩≥0\langle v,u\rangle=\langle v,qu\rangle\ge0 for all u∈Pu\in P; self-duality implies v∈Pqv\in P_q. Thus PqP_q is self-dual on qHqH.

Equations (13)–(14) give Jq(qMq)Jq=(qMq)′J_q(qMq)J_q=(qMq)'. Cone preservation restricts as well: if a∈pMpa\in pMp, then the operator θ(a)Jqθ(a)Jq\theta(a)J_q\theta(a)J_q on qHqH is the restriction of aJaJaJaJ. The center is qZ(M)qqZ(M)q, by (11) and (12), and JqcJq=c∗J_q cJ_q=c^* follows from MC. Hence the corner has all the displayed geometric properties of the constructed form. These conclusions correspond to Haagerup's Lemma 2.6, printed page 277, with the reduction facts here supplied by CG06.

CG08. A cyclic separating cone vector on a state corner

Suppose MM has a faithful normal state α\alpha. There is a cyclic separating vector γ∈P\gamma\in P.

Choose by the maximal principle a family of nonzero vectors in PP with mutually orthogonal algebra supports pip_i. If r=1−⋁ipi≠0r=1-\bigvee_i p_i\ne0, CG07 gives a nonzero projection rJrJrJrJ. Its cone is self-dual, and therefore cannot be {0}\{0\} on its nonzero Hilbert space. A nonzero vector in this cone has algebra support below rr, contradicting maximality. Thus ⋁ipi=1\bigvee_i p_i=1.

Faithfulness gives α(pi)>0\alpha(p_i)>0. Finite sums of their values are at most one. For each positive integer kk, only finitely many values can exceed 1/k1/k; their union shows that the family is countable. Normalize the vectors to have norm one and enumerate them as γi\gamma_i. Choose strictly positive scalars cic_i with ∑ici2<∞\sum_i c_i^2<\infty, and set

γ=∑iciγi∈P. \gamma=\sum_i c_i\gamma_i\in P.

Orthogonal supports make the series norm convergent. For b′∈M′b'\in M', each b′γib'\gamma_i lies in piHp_iH, so

∥b′γ∥2=∑ici2∥b′γi∥2. \|b'\gamma\|^2=\sum_i c_i^2\|b'\gamma_i\|^2.

If b′γ=0b'\gamma=0, it follows that b′γi=0b'\gamma_i=0 for every ii. Commutation with MM makes b′b' vanish on MγiM\gamma_i, and therefore on its closure qiHq_iH, where qi=JpiJq_i=Jp_iJ. The join of the qiq_i is one, so b′=0b'=0. The projection onto Mγ‾\overline{M\gamma} is consequently one. Since Jγ=γJ\gamma=\gamma, CG02 then also gives sγ=1s_\gamma=1. Thus γ\gamma is cyclic and separating. This proves the assertion without assuming a countable Hilbert-space dimension or a faithful normal state on the original arbitrary algebra.

CG09. Identifying the cone after choosing that vector

Let γ∈P\gamma\in P be cyclic and separating. The modular-core part of MC identifies its modular conjugation JγJ_\gamma with JJ, and hence identifies its natural cone with PP, as follows.

Let SS be the closure of aγ↦a∗γa\gamma\mapsto a^*\gamma. On this core the linear operator JSJS has nonnegative quadratic form:

⟨aγ,Ja∗γ⟩=⟨γ,a∗Ja∗Jγ⟩≥0.(15) \langle a\gamma,Ja^*\gamma\rangle =\langle\gamma,a^*Ja^*J\gamma\rangle\ge0. \tag{15}

It is therefore symmetric on the core, by polarization. The closed operator JSJS remains positive. Its adjoint is S∗JS^*J; the adjoint-core assertion in MC makes Mγ=JM′γM\gamma=JM'\gamma a core for that adjoint. On that core,

S∗J(aγ)=S∗((JaJ)γ)=(Ja∗J)γ=Ja∗γ. S^*J(a\gamma)=S^*((JaJ)\gamma) =(Ja^*J)\gamma=Ja^*\gamma.

Thus JSJS and its adjoint have the same core and action, so JSJS is positive self-adjoint. Its positive polar factor is (S∗S)1/2(S^*S)^{1/2}. The graph of the initial SS is invariant under interchanging its coordinates; its closure is too, and therefore S2u=uS^2u=u on its domain. This gives trivial kernel, while the range contains the dense space MγM\gamma. Uniqueness of polar decomposition now gives J=JγJ=J_\gamma.

MC now gives a self-dual cone PγP_\gamma generated by aJaJγaJaJ\gamma. It is contained in PP by cone preservation. Inclusion reverses under taking dual cones, so self-duality of both gives Pγ=PP_\gamma=P. This proves the comparison needed here directly; it does not use a comparison of two different weight representations.

CG09a. A dominated functional has a bounded commutant derivative

One elementary part of the analytic realization problem can be supplied without a representation theorem. Let γ\gamma be cyclic for MM, and let 0≤ψ≤Cωγ0\le\psi\le C\omega_\gamma. The rule

β(aγ,bγ)=ψ(a∗b)(16) \beta(a\gamma,b\gamma)=\psi(a^*b) \tag{16}

is a well-defined positive sesquilinear form on MγM\gamma, with

∣β(aγ,bγ)∣2≤ψ(a∗a)ψ(b∗b)≤C2∥aγ∥2∥bγ∥2. |\beta(a\gamma,b\gamma)|^2 \le\psi(a^*a)\psi(b^*b) \le C^2\|a\gamma\|^2\|b\gamma\|^2.

This both makes the definition independent of representatives and extends it continuously to HH. The Hilbert-space representation of bounded forms gives a unique 0≤T≤C10\le T\le C1 with β(u,v)=⟨u,Tv⟩\beta(u,v)=\langle u,Tv\rangle. For x∈Mx\in M, (16) gives

β(xaγ,bγ)=β(aγ,x∗bγ). \beta(xa\gamma,b\gamma)=\beta(a\gamma,x^*b\gamma).

Thus x∗T=Tx∗x^*T=Tx^* on a dense set and then on HH, so T∈M′T\in M'. Its bounded positive square root gives

ψ(a)=⟨γ,Taγ⟩=⟨T1/2γ,aT1/2γ⟩.(17) \psi(a)=\langle\gamma,Ta\gamma\rangle =\langle T^{1/2}\gamma,aT^{1/2}\gamma\rangle. \tag{17}

This proves bounded commutant realization for a dominated functional. It does not assert that T1/2γT^{1/2}\gamma lies in the natural cone, and so does not replace CR.

CG10. Positive functionals on an arbitrary algebra

Assume MC and CR. Every ω∈M∗+\omega\in M_*^+ has a unique vector ξω∈P\xi_\omega\in P with

ω(a)=⟨ξω,aξω⟩(a∈M).(18) \omega(a)=\langle\xi_\omega,a\xi_\omega\rangle \quad(a\in M). \tag{18}

No faithfulness, separability, countable decomposability, or cardinality restriction on MM is imposed.

If ω=0\omega=0, use zero. Otherwise let p=s(ω)p=s(\omega), q=pJpJq=pJpJ, and use (12) to transport ω∣pMp\omega|_{pMp} to the von Neumann algebra qMqqMq. By CG01 this transported functional is faithful and normal, and its normalization is a faithful normal state. CG07–CG08 give a cyclic separating cone vector on qHqH; CG09 identifies its natural cone with PqP_q. CR gives a vector ξ∈Pq\xi\in P_q representing the transported functional. For every a∈Ma\in M,

⟨ξ,aξ⟩=⟨ξ,qaqξ⟩=ω(pap)=ω(a), \langle\xi,a\xi\rangle =\langle\xi,qaq\xi\rangle =\omega(pap)=\omega(a),

where the last equality is (1). The vector lies in PP, and uniqueness follows from CG05. This closes the arbitrary-cardinality and nonfaithful reduction. It is Haagerup's Lemma 2.10 reduction, printed pages 278–279, with the corner algebra and cyclic cone identification proved above.

Combining CG01, CG02, CG05 and (18) gives precisely

∥ξω−ξψ∥2≤∥ω−ψ∥,qξω=Js(ω)J,ψ≤ω⟹s(ψ)≤s(ω),∥ξω∥2=ω(1).\begin{gathered} \|\xi_\omega-\xi_\psi\|^2\le\|\omega-\psi\|,\qquad q_{\xi_\omega}=Js(\omega)J,\\ \psi\le\omega\Longrightarrow s(\psi)\le s(\omega),\qquad \|\xi_\omega\|^2=\omega(1). \end{gathered}

The proof of these geometric formulas is complete above. Existence of the initially constructed self-dual cone remains the MC construction input; existence in the cyclic natural cone remains CR. The cyclic input is Araki's Theorem 6, printed pages 335–339; its analytic construction uses positive affiliated-operator approximation in the cyclic representation.

References

Huzihiro Araki, Some properties of modular conjugation operator of von Neumann algebras and a non-commutative Radon–Nikodym theorem with a chain rule, Pacific Journal of Mathematics 50 (1974), 309–354. Theorem 4(6)–(8) gives the orthogonal cone decomposition, support and squared-distance conclusions; Theorem 6 gives the cyclic normal-functional construction.

Uffe Haagerup, The standard form of von Neumann algebras, Mathematica Scandinavica 37 (1975), 271–283. Lemma 2.6 treats the standard corner and Lemma 2.10 passes from cyclic normal-functional representatives to arbitrary algebras. The corner commutant, support and norm arguments used in that passage are proved above.